Apply machine learning methods to diagnose colorectal cancer (crc) Áp dụng phương pháp học máy để chẩn đoán ung thư đại trực tràng (CRC)
Overview of the Colon
Overview of the Colon
The colon is a crucial component of the human digestive system, comprising the final sections of the large intestine, which connects the small intestine to the rectum Measuring approximately 1.5 to 1.8 meters in length, the colon and rectum play vital roles in processing, storing, and excreting waste from the body.
The colon is the largest and longest part of the large intestine, connecting from the ileum (the last part of the small intestine) to the rectum
The colon consists of five main sections: the cecum, ascending colon, transverse colon, descending colon, and sigmoid colon, each of which is essential for the absorption of water and electrolytes, as well as the formation of stool.
Figure 1: Colon in the human body
The structure of the colon wall generally has 4 layers:
- The serosa layer: created by the visceral layer of the peritoneum has an omental diverticulum
The subserosal layer of the colon consists of an outer layer of longitudinal muscle organized into three distinct bands, with a thin section between these bands The inner layer features the sphincter, while the Auerbach neural network is situated between the circular and longitudinal muscle layers, playing a crucial role in gastrointestinal function.
- Submucosal layer: is a connective tissue with many blood vessels and nerves Located in the submucosa is the Meissner neural network [5]
The colon's mucosal layer is distinct from that of the small intestine, as it lacks villi and features a flat surface with numerous straight tubular cavities These cavities extend into the stromal layer and are lined with mucous secretory cells, along with occasional endocrine cells Additionally, the mucosal layer contains lymphoid tissue, contributing to the colon's immune function.
Effects of the colon on the human body
The colon's primary role is to process digested food by absorbing water and electrolytes, while also breaking down waste with the help of bacteria It compacts this waste into feces and stores it until there is sufficient volume Once ready, the colon contracts to facilitate peristalsis, leading to the excretion of stool through the rectum, the final segment of the colon located near the anus.
The large intestine hosts a variety of bacteria that play a crucial role in protein synthesis and the production of essential vitamins, including vitamin K, vitamin B12, thiamin, and riboflavin These bacteria also contribute to gas production within the intestine Notably, vitamin K is vital for maintaining proper blood clotting, as dietary sources alone are often inadequate to meet the body's needs.
The colon plays a vital role in digestion by maintaining an alkaline environment that aids in the further breakdown of undigested food remnants, which have already been processed in the acidic conditions of the stomach and small intestine Additionally, the mucosa of the colon secretes a small amount of alkaline fluid, which serves to protect the intestinal lining and help soften stools for easier passage.
The colon mucosa is primarily composed of mucous cells that secrete mucus, which is released when food interacts with these cells or when they are activated by local enteric reflexes This mucus serves to protect the intestinal wall from abrasions and the harmful effects of bacteria present in stool, while also helping to bind the stool together.
In addition, the digestive tract is also the place where the body's residues and drugs are excreted after ingestion
The absorption function primarily takes place in the first half of the large intestine, where the mucosa exhibits a significant absorptive capacity This process involves the continued absorption of water from the small intestine, which is then concentrated in the kidneys and transformed into waste for fecal excretion Additionally, the colon plays a crucial role in absorbing mineral water and other essential elements.
Reduced colon movement can lead to prolonged retention of waste in the intestines, resulting in the absorption of excess water and the formation of dry, hard stools, which causes constipation Conversely, when fecal matter moves too quickly through the colon, such as in cases of acute enteritis, the colon mucosa secretes significant amounts of water and electrolytes to dilute irritants and expedite stool passage, leading to diarrhea While diarrhea can cause dehydration and electrolyte loss, it also effectively eliminates irritants from the body, aiding in the patient's recovery.
Description of the research problem
Causes and Symptoms of Colorectal Cancer
Colorectal cancer is on the rise, particularly among younger individuals, as it develops in the rectum or colon and often goes unnoticed until the condition worsens It's crucial to take even minor symptoms seriously, as early detection is key Risk factors for colon cancer include age, a history of colon polyps, inflammatory bowel disease, and family history of the disease Individuals experiencing symptoms such as constipation, diarrhea, blood in stools, abdominal pain, or changes in bowel habits should be vigilant, as these may indicate an increased risk of colorectal cancer Key symptoms to watch for include vomiting, poor appetite, and persistent gastrointestinal issues.
Recent weight loss has been observed in patients who were previously skinny, accompanied by symptoms such as diarrhea—sometimes bloody—and yellowing of mucous membranes Additional signs include skin darkening, swollen peripheral and supraclavicular lymph nodes, and abdominal distention Clinical symptoms may vary, presenting as vomiting, diarrhea, dehydration, irritability, and lethargy, while non-clinical symptoms often consist of weight loss, bloating, poor appetite, and decreased milk production To determine the potential for infection, it is crucial for patients to undergo regular health checks.
Colon cancer in the world and Vietnam
Colorectal cancer ranks as the fourth most common cancer among men and the third among women globally, with significant variations in incidence rates across different regions According to Meara's data from 1998-2002, rates per 100,000 men ranged from as low as 4 in India (Karunagappally) to 59.1 in the Czech Republic For women, the consumption of refined grains, fruits, and vegetables varies between 3 to 5 servings daily The lowest incidence rate was reported in Karunagappally, India, at 6, while New Zealand recorded a much higher rate of 39.5 High incidence rates are predominantly found in Europe, North America, and Oceania, where death rates are nearly double those in South-Central Asia, with the lowest rates observed in Asian, African, and Latin American regions.
Colorectal cancer is a significant health concern in Vietnam, ranking 5th among all cancer types, following stomach, lung, breast, and palate cancers Central K Hospital reports that this type of cancer accounts for 9% of the total cancer cases in the country According to Associate Professor Tran Van Thuan, the Director of K Hospital, colorectal cancer is the 4th most common cancer among men, with over 8,000 new cases diagnosed annually Overall, Vietnam sees approximately 16,000 new colorectal cancer cases each year.
Colorectal cancer rates are rising both nationally and globally, largely due to the challenge of early symptom detection and the necessity for regular screenings Consequently, recognizing the early signs of this disease is crucial for decreasing the incidence of colon cancer.
Objectives of the analysis
The goal is to analyze and point out common symptoms in colorectal cancer patients Identifying important symptoms will enhance early diagnosis and timely treatment, thereby improving patient survival rates
Utilizing machine learning techniques to develop diagnostic models from medical symptoms enables clinicians to swiftly identify high-risk cases and recommend essential follow-up tests This innovative tool enhances diagnostic precision while simultaneously lowering both medical expenses and time spent on patient care.
To reduce the risk of disease, it is essential to adopt healthy lifestyle changes and engage in regular exercise Additionally, periodic screening tests, such as colonoscopies, are recommended to detect and promptly address pre-cancerous lesions, ensuring early intervention and better health outcomes.
Compared with previous studies in the same field, such as the study by Smith et al
Recent studies, including Johnson et al (2022), have highlighted the effectiveness of ensemble methods, particularly the Stacking model, in optimizing predictions for colorectal cancer outcomes Unlike previous research that relied on single models or simpler techniques to address imbalanced data, the Stacking approach combines multiple powerful models, significantly enhancing prediction accuracy and improving the classification of positive cases This advancement is crucial in the medical field, as early detection of colorectal cancer can save lives and reduce treatment costs.
Our research enhances the diagnosis and treatment process by integrating machine learning models, which improve diagnostic quality while reducing costs and time These models serve as effective clinical decision support tools within health systems, enabling physicians and medical professionals to make more accurate and timely decisions.
To enhance the efficiency and applicability of models in the future, significant advancements are essential Emphasizing model optimization is crucial for achieving superior performance and improved explainability, particularly in the medical field, where clarity and transparency in decision-making are vital.
Developing interpretation methods such as SHAP values or LIME will help increase medical professionals' trust and acceptance of machine learning systems
Integrating various types of biomedical data is a crucial area of research Utilizing advanced techniques like deep learning and transfer learning can enhance the analysis and prediction capabilities of large, complex datasets.
Ensuring the security and privacy of medical data is crucial Implementing robust data security measures, such as encryption and secure access protocols, will safeguard sensitive patient information Additionally, exploring and utilizing privacy protection techniques like differential privacy and federated learning during model training and usage is essential for maintaining confidentiality.
To ensure the widespread adoption of machine learning models in healthcare, it is essential to create seamless integration solutions Developing user-friendly interfaces and clinical decision support systems will enable healthcare professionals to effectively utilize these models in diagnosis and treatment Furthermore, ongoing research and testing of advanced learning models, such as Generative Adversarial Networks (GANs) and Reinforcement Learning, will unlock new possibilities for applying machine learning in medicine and enhancing predictive capabilities.
This research has yielded significant results and valuable contributions to medical diagnosis Ongoing optimization and study of these models are essential for enhancing performance, security, and practical integration, ultimately leading to improved healthcare quality and better patient outcomes.
Research Tools
Python Language
Python, created by Guido van Rossum and released in 1991, is a highly flexible, interpreted, high-level programming language known for its easy-to-read and easy-to-learn syntax With a strong user community, Python has become the most popular programming language in 2024, according to the official Python Developer Twitter page It is widely utilized in machine learning, significantly contributing to research in colorectal cancer.
Google Colab
Google Colaboratory, or Google Colab, is a free web-based programming environment by Google that enables users to write and execute Python code in their browser It is particularly beneficial for tasks involving machine learning, data analysis, and AI model development In colorectal cancer research, Google Colab serves as a key platform, enhancing machine learning applications throughout the research process.
Machine Learning
Machine learning, a subset of artificial intelligence (AI), utilizes algorithms trained on data sets to create autonomous models that predict and categorize information independently Its diverse applications include product recommendations based on consumer behavior, stock market trend forecasting, and language translation services.
"Machine learning" and "artificial intelligence" are frequently confused, but they have different meanings Artificial intelligence refers to the overarching aim of equipping machines with human-like cognitive skills, whereas machine learning focuses specifically on using algorithms and data sets to achieve this goal.
METHODOLOGY
Preprocessing Data
Data collection serves as the cornerstone of our analysis, as evidenced by our comprehensive dataset gathered from 1,755 colorectal cancer patients at Thai Nguyen Hospital This dataset encompasses 21 symptoms along with their diagnostic results, which are crucial for developing accurate predictive models.
Before analyzing data, it must be cleaned and prepared to ensure accuracy and consistency Data preprocessing includes several steps:
Normalization: Adjusting data values to a common scale, usually from 0 to 1, to ensure uniformity within a data set
Clean: Resolves missing values and fixes errors We consulted with medical experts on unrecorded diagnoses, ensuring that all data points were usable
Encoding transforms categorical symptoms into binary or numeric formats, enabling their use in machine learning models This crucial step ensures that raw medical data is converted into a format that models can process efficiently, enhancing the overall effectiveness of data analysis in healthcare.
To effectively analyze the data, it is essential to examine the relationship and correlation between symptoms through the use of correlation matrices and column charts These visual tools enhance the understanding of data distribution and facilitate the development of an appropriate machine learning method In this analysis, the exploratory data analysis (EDA) techniques employed include correlation matrices and bar charts.
A correlation matrix is a comprehensive table that displays correlation coefficients between various variables, effectively summarizing large data sets and revealing patterns The correlation coefficient, which ranges from -1 to 1, indicates the strength and direction of the relationship between pairs of values; -1 signifies a perfect negative correlation, +1 indicates a perfect positive correlation, and 0 denotes no correlation In this context, the matrix cells illustrate the correlation degree between pairs of macrophage cancer symptoms.
Bar plots effectively visualize the distribution and frequency of various symptoms within a dataset by utilizing bar charts Each symptom is represented by a separate bar chart, allowing for a clear representation of the values associated with these symptoms This visual approach aids in comprehending the data and its characteristics, while also facilitating the identification of trends or anomalies present in the dataset.
Handling imbalanced data
To solve the imbalance problem, we have two quite common methods: Oversampling and Undersampling
Oversampling is a method used to enhance the representation of minority class samples in imbalanced datasets, aiming to balance the class ratio This technique helps prevent the model from being biased towards the majority class, allowing it to effectively learn and recognize features from the minority class.
Undersampling is a method used to address class imbalance in datasets by decreasing the number of samples from the majority class This technique aims to balance the class distribution, allowing the model to learn effectively from both classes By mitigating the influence of the majority class, undersampling enhances the model's ability to identify features from all classes equally.
Figure 4: Distinguishing Undersampling and Oversampling
Oversampling is an effective technique for balancing classes in datasets, as it retains all original data while generating additional samples This method enhances the model's ability to understand and predict minority classes, ultimately improving performance Unlike undersampling, which risks significant data loss and potential overfitting due to a reduced dataset, oversampling maintains data integrity and supports robust model training.
The two main Oversampling methods used in this problem are SMOTE and Upsampling
SMOTE enhances the representation of the minority class in a dataset by randomly selecting a data point and identifying its nearest neighbors It generates new samples through interpolation between the original point and these neighbors, repeating the process until the desired sample size is achieved Unlike simple duplication of existing data, SMOTE produces meaningful new samples, which significantly boosts the model's predictive capabilities for the minority class This approach enables the model to learn from a more diverse set of examples, ultimately improving its generalization ability.
Upsampling is a method performed by copying the existing samples of the minority class many times until this class reaches the same size as the majority class
Replicated samples are added to the data set, combined with the majority class samples to create a more balanced data set.
Model Selection and Training
Supervised and unsupervised machine learning are among the most prevalent types, each serving distinct purposes in model training and data analysis
Supervised and unsupervised machine learning models are two main subgroups of machine learning algorithms, depending on the nature of the training data and the goal of the learning
Supervised learning is a machine learning approach that utilizes labeled datasets to train algorithms for accurate data classification and outcome prediction By employing both input and output data, this method enables the model to assess prediction accuracy and improve its performance over time.
Supervised learning can be divided into two types of data mining problems: classification and regression
In classification problems, algorithms accurately categorize test data, such as differentiating between rice grains and beans In the medical field, supervised learning algorithms classify X-rays as normal, benign, or malignant Popular classification algorithms include linear classification, support vector machines, decision trees, and random forests.
Regression is a key supervised learning technique that employs algorithms to analyze the relationship between dependent and independent variables It is primarily used to forecast numerical values by interpreting various data points Notable regression algorithms include linear regression, logistic regression, and polynomial regression.
Unsupervised machine learning models use algorithms to analyze and cluster unlabeled data sets These algorithms discover hidden patterns in data without human intervention, hence they are called "unsupervised"[17]
Unsupervised learning models are applied to three main tasks: clustering, association, and dimensionality reduction:
Clustering is a data mining technique that organizes unlabeled data points by identifying their similarities or differences A prominent example is the K-means clustering algorithm, which categorizes similar data points into distinct groups, with the K value indicating the number of groups and the granularity of the analysis This approach is particularly beneficial in the medical field, where clustering algorithms can effectively classify patients based on clinical signs and test results.
Association is a form of unsupervised learning that identifies relationships between variables in a dataset In the financial sector, these techniques are frequently employed for anomaly detection, allowing for the identification of fraudulent transactions and unusual financial activities.
Dimensionality reduction is a crucial technique employed to manage large datasets with numerous features by condensing the input data while preserving its integrity Commonly utilized in the data preprocessing phase, this method helps eliminate noise from image data, thereby enhancing overall image quality.
Logistic regression, also known as the logit model, is a powerful tool for data classification and predictive analysis in machine learning As a supervised machine learning model, it functions as a discriminative model, focusing on distinguishing between different classes or categories By estimating the probability of specific events, such as voting behavior, logistic regression utilizes a dataset of independent variables for its predictions There are three main types of logistic regression: binary logistic regression, multinomial logistic regression, and ordinary logistic regression.
The logistic function is an S-shaped curve that models the probability of a sample belonging to a specific class It ensures that the dependent variable remains within the range of 0 to 1, irrespective of the independent variable's value.
Figure 7: S-shaped plot of Logistic regression function
Logistic regression is a statistical model that uses the logistic function, or logit function in mathematics, as the equation between x and y The logit function maps y as a sigmoid function of x
In numerous instances, various explanatory variables affect the dependent variable's value To effectively model these datasets, the logistic regression formula requires the assumption of a linear relationship among the independent variables.
Binary logistic regression is an ideal model for analyzing binary data, where the dependent variable has two possible outcomes, such as symptomatic or asymptomatic This method is widely used for binary classification and is particularly effective for making accurate predictions in such scenarios.
Random Forest is a versatile machine learning algorithm that aggregates the results of various decision trees to produce a unified outcome, a technique referred to as method binning (GeeksforGeeks, 2024) Its user-friendly nature and adaptability have led to widespread use, as it effectively addresses both classification and regression challenges while being closely associated with Decision Trees and Ensemble methods.
A decision tree is a supervised learning algorithm characterized by its hierarchical data structure It is built by utilizing a random subset of the dataset, which allows for the evaluation of a random selection of features within each partition.
The random forest model consists of multiple decision trees, making it essential to understand the decision tree algorithm first A decision tree begins with a fundamental question, leading to a sequence of inquiries that help in reaching a conclusion.
The root decision represents the primary question of a problem, while the branches connect this root to related questions that lead to the final answer Each node corresponds to the answer of the immediate branch question After addressing all questions, the responses are compiled, and the most frequently occurring answer is selected as the decision for the original question.
Random forests are valuable for assessing the significance of variables in predicting disease symptoms, as they minimize reliance on specific sample data and lower the likelihood of overfitting.
Evaluate models
This article discusses the evaluation of model performance through two primary methods: Model Performance Analysis and Confusion Matrix The Model Performance Analysis method employs metrics such as Accuracy, Precision, Recall, F1 Score, and AUC to assess the effectiveness of the model.
Accuracy is the ratio of the total number of correct predictions to the total number of predictions
Precision is the ratio of correct Positive predictions compared to the total number of Positive predictions
Recall is the ratio of correct Positive predictions to the total number of actual Positives
The F1 Score is the harmonic average of Precision and Recall, indicating the balance between them
AUC is the area under the ROC curve, which measures a model's ability to discriminate between classes
The Confusion Matrix is a crucial tool for evaluating the performance of classification models, providing a clear and structured way to assess their accuracy on datasets with known true values This matrix comprises four key components that help in understanding the effectiveness of the model's predictions.
Positives (TP): Number of positive samples that were correctly predicted to be positive
True Negatives (TN): Number of positive negative samples correctly predicted as positive negatives
False Positives (FP), often called "Type I Error": The number of negative positive samples that were mistakenly predicted as positive positives
False Negatives (FN), often called "Type II Error": The number of positive samples that were mistakenly predicted as positive negatives
From the Confusion Matrix, calculating many other evaluation indexes such as Accuracy, Recall, Precision, F1 score, and AUC also helps evaluate the model better with
Utilizing a combination of two methods not only enhances the identification of the most effective model but also ensures the selection of an accurate prediction model, thereby reducing the risk of bias towards a single outcome.
There are also other model evaluation methods used such as:
The ROC Curve is an essential tool for assessing the effectiveness of a classification model, illustrating the relationship between the True Positive Rate (TPR) and the False Positive Rate (FPR) across various classification thresholds TPR measures the proportion of actual positives accurately identified by the model, while FPR indicates the rate of false positives.
(FPR) is the proportion of true negatives mistakenly predicted as positive by the model count
Precision-Recall Curve represents the relationship between Precision (accuracy of positive prediction) and Recall (rate of correctly identifying positive cases) at different classification thresholds
The optimal threshold value, known as the best Threshold, is where the F-Score achieves its maximum The F-Score serves as a comprehensive metric that balances both Precision and Recall, offering an integrated perspective on these two important performance indicators.
DATA PREPROCESSING AND MODEL
Data sources and key variables
Data is a data set of 1755 colorectal cancer patients (1755 data rows) of Thai Nguyen Hospital with 22 columns corresponding to 21 symptoms that patients have and 1 KQ column (Results)
The 22 symptoms include: abdominal pain, vomiting, anorexia, constipation, weight loss, diarrhea, bloody stools, yellow skin and mucous membranes, dark skin, peripheral lymph nodes, supraclavicular lymph nodes, abdominal distention, and side effects abdomen, peritoneal tenderness, crawling snake sign, floating bowel loops, palpable tumor, rectal examination with tumor, history of cancer, abdominal CT scan with tumor, colonoscopy with tumor
The KQ column is the diagnostic conclusion from those 22 symptoms
The data set is represented as a binary of 0 and 1 0 is “Asymptomatic” and 1 is
Figure 12: The dataset of the research article
Data preprocessing Data preprocessing includes
Standardized "diagnosis" data on "Cancer" and "Non-cancer", in some cases without a diagnosis, a specialist's opinion is required
Normalize extra words and whitespace characters in the data of "First day developments"
"First-day developments" data separates typical symptoms according to expert opinion, removing unrelated symptoms
Standardize disease symptoms as suggested by doctors and pathologists on rectal cancer
Ask an Oncologist for additional diagnosis for patients who do not have a diagnosis Filter out typical symptoms related to rectal cancer from specialized oncologists
Table 1: Example of data normalization results
CHẨN ĐOÁN TRIỆU CHỨNG DIỄN BIẾN NGÀY ĐẦU TIÊN ung thư nội soi có ổ đại tràng, da - xanh nhẹ, da niêm mạc - hồng nhạt, bụng - mềm, chán ăn
Bệnh nhân vào viện với chẩn đoán u ác của đại tràng góc gan, đã được xác định là ung thư biểu mô tuyến với di căn gan đa ổ và di căn phúc mạc Sau khi điều trị hóa chất mFOLFOX6 đợt 1, tình trạng bệnh nhân ổn định nhưng vẫn có dấu hiệu mệt mỏi, ăn uống kém, và phù hai chi dưới Khám lâm sàng cho thấy bệnh nhân tỉnh, da hơi xanh, không xuất huyết, và các cơ quan khác bình thường Tiền sử bệnh nhân có tăng huyết áp, đã trải qua hai lần tai biến mạch máu não và phẫu thuật cắt 2/3 dạ dày Chẩn đoán hiện tại bao gồm u ác của đại tràng góc gan, theo dõi thiếu máu, thiếu protein năng lượng, cao huyết áp vô căn, và nghi ngờ Covid-19, chờ kết quả xét nghiệm để có hướng xử trí tiếp theo.
Bệnh nhân vào viện do sốt cao (Tmax 39,5°C) và rét run, với tiền sử ung thư đại tràng, gan và phổi Sau khi điều trị tại khoa ung bướu, bệnh nhân xuất hiện các triệu chứng như đau đầu, mỏi toàn thân, mệt mỏi, ăn uống kém, ho có đờm trắng, đau tức ngực và khó thở nhẹ Bệnh nhân cũng gặp tình trạng táo bón bình thường và không thấy cải thiện sau khi dùng thuốc Lúc vào viện, bệnh nhân tỉnh táo nhưng có sốt cao.
Bệnh nhân có triệu chứng đau đầu, mệt mỏi, đau tức ngực và khó thở nhẹ Xuất hiện đau bụng thượng vị và có dấu hiệu nhiễm trùng với da mệt, lưỡi có nốt đỏ, mủ, và thâm Niêm mạc không phù, không có xuất huyết tại vị trí tiêm truyền, hạch ngoại vi không to Tim nhịp nhanh đều với tần số 100 lần/phút, phổi khó thở nhưng không có rales Bụng chướng và gan lách có dấu hiệu bất thường Thân nhiệt đạt 39,4°C, huyết áp 100/60 mmHg Chẩn đoán sơ bộ cho thấy bệnh nhân có thể mắc sốt CRNN và có khả năng liên quan đến bệnh lý đại tràng, không có dấu hiệu ung thư da niêm mạc.
Tiền sử: THA không điều trị đều, PT cắt đại tràng
Năm 2018, bệnh nhân sau nhiều lần phẫu thuật gặp phải tình trạng đau bụng âm ỉ quanh rốn, thỉnh thoảng xuất hiện cơn quặn và buồn nôn, đã nhập viện mà chưa được xử trí Khám lâm sàng cho thấy bệnh nhân tỉnh táo, tiếp xúc tốt, da niêm mạc hồng hào, không có triệu chứng nôn, sốt hay khó thở, nhưng có đau âm ỉ quanh rốn và bụng chướng nhẹ Bệnh nhân có dấu hiệu đau nhiều quanh rốn và phản ứng thành bụng âm tính Siêu âm cho thấy hình ảnh theo dõi gan, chẩn đoán ban đầu là tắc ruột sau mổ Bệnh nhân không có tiền sử ung thư, tình trạng da niêm mạc bình thường, và các dấu hiệu lâm sàng như đau bụng quanh rốn và dấu hiệu rắn bò đã được ghi nhận.
Bệnh nhân nữ 57 tuổi nhập viện do đau bụng, có tiền sử mổ cắt u đại tràng cách đây 1 tháng Bệnh nhân không có tiền sử tăng huyết áp, đái tháo đường, và không bị tiêu chảy Triệu chứng đau bụng quanh rốn xuất hiện theo từng cơn, kèm theo buồn nôn nhưng không nôn Khi khám, bệnh nhân tỉnh táo, da và niêm mạc hồng, không sốt hay khó thở, với mạch 90 lần/phút và huyết áp 150/80 Bụng bệnh nhân chướng, có phản ứng thành bụng và dấu hiệu rắn bò dương tính, trong khi cảm ứng phúc mạc âm tính Chẩn đoán sơ bộ là tắc ruột sau mổ cắt u đại tràng, với triệu chứng đau bụng hạ vị, táo bón, bụng chướng nhẹ, da niêm mạc nhạt và chán ăn.
Bệnh nhân có tiền sử phẫu thuật cắt u ruột non vào tháng 4/2020 và cắt đoạn đại tràng do u vào tháng 6/2021 Hiện tại, bệnh nhân gặp phải triệu chứng đau bụng quanh rốn và hạ vị từng cơn, kèm theo buồn nôn và tình trạng đi tiêu bất thường Khi vào viện, bệnh nhân tỉnh táo nhưng thể trạng gầy yếu, da niêm mạc nhạt và có dấu hiệu suy kiệt Bụng chướng nhẹ, không có dấu hiệu của bụng ngoại khoa, tim nhịp đều và phổi thông khí rõ Siêu âm cho thấy gan có kích thước bình thường nhưng nhu mô có nhiều khối giảm âm.
35 xung quanh có viền giảm âm, kích thước lớn nhất 13 x 18 mm Chẩn đoán: Tắc ruột do K đại tràng M gan Handling noisy data:
Statistics on the frequency of symptoms from Oncology Department data
Table 2: Symptom statistics and numbers
STT Thuộc tính Ung thư Không ung thư Tổng
3 da niêm mạc - hồng nhợt
4 hạch ngoại vi - âm tính 497 3 500
5 hạch ngoại vi - không to 372 8 380
12 da niêm mạc - bình thường 129 5 134
17 hạch ngoại biên - âm tính 81 1 82
20 nội soi đại tràng có u 66 2 68
23 da niêm mạc - nhợt (nhạt) 58 6 64
26 da niêm mạc - nhợt nhẹ 56 0 56
28 chụp ct ổ bụng có khối u 47 2 49
40 cảm ứng phúc mạc - âm tính 17 2 19
Isolate 288 symptoms that satisfy the symptoms as suggested by the doctor
Meet and consult with specialized doctors to combine and eliminate symptoms to help reduce diversity in describing symptoms
Table 3: Pooled table of colorectal cancer attributes
1 bung chuong/chuong bung/bung cứng
2 bung mềm/bung không trưởng
3 cảm ứng phúc mạc - âm tính / cảm ứng phúc mạc (-)
4 cảm ứng phúc mạc/cảm ứng phúc mạc - dương tính/cảm ứng phúc mạc - nghi ngờ
5 chụp ct ổ bụng có khối u/ct ở đại tràng sigma/ct ổ bụng có khối u/có khối ở hố chậu phải/chụp ct ổ bụng có hình ảnh ung thư di căn
6 da sạm/da sạm - rất đau
7 da vàng/da vàng sạm/da niêm mạc - vàng/da niêm mạc - vàng nhạt
8 da xanh/da xanh nhẹ/da xanh tím/da hơi xanh
10 đau bụng/đau bụng - âm i/đau bụng - âm i hạ vị/đau bụng - âm i liên tục/đau bụng
Đau bụng có thể xuất hiện ở nhiều vị trí khác nhau, bao gồm âm i quanh rốn, thượng vị, vùng gan, hố chậu trái, và hạ sườn Cảm giác đau có thể là âm i hoặc tức, tùy thuộc vào nguyên nhân Đặc biệt, đau ở hạ sườn phải và hạ vị cũng cần được chú ý, vì đây có thể là dấu hiệu của các vấn đề sức khỏe nghiêm trọng Việc xác định chính xác vị trí đau bụng là rất quan trọng để chẩn đoán và điều trị hiệu quả.
Đau bụng có thể xuất hiện ở nhiều vị trí khác nhau như hạ vị, thượng vị, và quanh rốn Các triệu chứng đau bụng thường được mô tả là chắc, nặng hoặc nhẹ, và có thể kéo dài liên tục hoặc từng cơn Đau có thể lan tỏa đến hố chậu trái, mạn sườn phải, hoặc vùng gan Tùy thuộc vào vị trí và mức độ đau, người bệnh có thể cảm thấy đau quanh rốn, đau hạ vị, hoặc đau vùng hạ sườn Những cơn đau này có thể kèm theo cảm giác tức, nặng hoặc nhẹ và có thể liên quan đến các vấn đề tiêu hóa hoặc các bệnh lý khác Việc xác định chính xác vị trí và tính chất của cơn đau bụng là quan trọng để tìm ra nguyên nhân và phương pháp điều trị phù hợp.
Đau bụng có thể xuất hiện ở nhiều vị trí khác nhau như hạ vị, hố chậu phải, quanh rốn và hố chậu trái Các cơn đau này có thể là đau âm ỉ, đau tức, hoặc đau nhẹ, đau nhiều Đặc biệt, đau vùng thượng vị và hạ vị thường kèm theo cảm giác khó chịu Đau có thể tỏa ra từ vùng gan, hạ sườn trái và phải, gây ra cảm giác đau tức hoặc đau âm ỉ Những triệu chứng này cần được chú ý, vì chúng có thể liên quan đến các vấn đề sức khỏe nghiêm trọng hơn.
11 dấu hiệu quai rắn bo/dấu hiệu rắn bo/dấu hiệu rắn bo - dương tính
Hạch thượng đòn là những hạch bạch huyết quan trọng, có thể xuất hiện dưới nhiều dạng khác nhau Các loại hạch thượng đòn bao gồm: hạch thượng đòn trái, hạch thượng đòn chắc, hạch thượng đòn di động hạn chế, và hạch thượng đòn dính Ngoài ra, có thể thấy hạch xương đòn di động và nhiều hạch nhỏ xung quanh Việc nhận diện và phân loại những hạch này là cần thiết để đánh giá tình trạng sức khỏe và phát hiện các vấn đề tiềm ẩn.
Hạch ngoại biên không phát hiện, không sờ thấy và không to, bao gồm cả hạch ngoại vi tuyến giáp, đều có kích thước nhỏ và âm tính.
14 niêm mạc (da niêm mạc) - bình thường/hồng/hồng nhợt/hồng nhạt/nhợt/nhợt nhẹ/tái nhợt
15 nội soi có ổ đại tràng/nội soi có ổ manh tràng/nội soi đại tràng có ổ/nội soi manh tràng có khối u/nội soi trực tràng có khối u
16 phân - có máu/phân - có máu đỏ tươi/phân - đen/phân - nhỏ/phân vàng nâu/táo bón/tiêu chảy
17 phản ứng thành bụng - âm tính
18 phản ứng thành bụng/phản ứng thành bụng - dương tính/phản ứng thành bụng - nghi ngờ
19 quai ruột nổi - âm tính/dấu hiệu quai ruột nổi - âm tính/dấu hiệu rắn bo - không rõ
20 quai ruột nổi/quai ruột nổi - dương tính/quai ruột giãn
21 rối loạn đại tiện/rối loạn tiêu hóa
Khối u ở ổ bụng có thể được phát hiện qua siêu âm hoặc khi sờ thấy Các loại khối u thường gặp bao gồm khối u đại tràng và u sùi đại tràng phải, có thể dẫn đến tình trạng tắc ruột Việc thăm khám đại tràng và trực tràng là cần thiết để xác định sự hiện diện của khối u, đặc biệt là trong trường hợp có tiền sử nội soi đại tràng.
23 sút cân/thể trạng - gầy/thể trạng - gầy suy kiệt/thể trạng yếu/thể trạng suy kiệt/thể trạng trung bình
The data were pooled from 288 to 24 major symptoms
After performing the second assessment, eliminating additional confounding symptoms, the remaining 21 symptoms include: "abdominal pain", "vomiting", "loss of appetite", "constipation", "weight loss", "diarrhea" bleeding”, “bloody stools”,
“yellow skin and mucous membranes”, “dark skin”, “peripheral lymph nodes”,
“supraclavicular lymph nodes”, “abdominal distention”, “abdominal wall reaction”,
“peritoneal tenderness ", "snake crawling sign", "floating intestinal loop", "palpable
38 tumor", "rectal examination shows a tumor", "cancer history", "abdominal CT scan shows a tumor", "internal Colonoscopy showed tumor"
Data transformation: Transform data as input to the decision tree model
The data table consists of 21 columns detailing symptoms alongside cancer diagnosis results By encoding the input with a value of 0.1 in the decision tree, it indicates the presence or absence of patient symptoms observed during the preclinical examination, ultimately leading to a final diagnosis of whether the patient has cancer or not.
EDA
Research indicates a strong association between abdominal bloating and severe abdominal pain in colorectal cancer patients, with a correlation coefficient of 0.67 This suggests a potential common pathological mechanism or disease progression pattern Additionally, flatulence is also linked to these symptoms, highlighting the interconnected nature of gastrointestinal issues in this patient population.
Irritation, indicated by peroneal tenderness, may signal an inflammatory response or tumor presence in the abdominal cavity Additionally, a slight negative correlation between bloating and diarrhea suggests that as patients experience increased bloating, they may encounter reduced diarrhea, potentially reflecting different manifestations or stages of the disease.
Symptoms that exhibit correlation values near 1 or -1 with "KQ" suggest a strong relationship; however, none of the symptoms reached an extremely high correlation This indicates that no single symptom is particularly significant in predicting "KQ."
"KQ" but that "KQ" may depend on a combination of many symptoms together
The correlation matrix is an essential tool for analyzing the relationships between various symptoms and their impact on the outcome variable of colorectal cancer, offering a foundation for enhancing diagnostic accuracy.
Figure 14: Correlation between independent variables
Bar plots effectively visualize the distribution and frequency of various symptoms within a dataset by utilizing bar charts Each symptom is represented in a separate chart, allowing for a clearer understanding of the data and the unique characteristics associated with each symptom This visualization aids in identifying trends or anomalies, ultimately highlighting significant symptoms that warrant attention.
There were 157 cases of vomiting and 1598 cases without symptoms On average, about 9% of people experience vomiting
There were 104 cases of abdominal distention and 1651 cases without symptoms On average, about 6% of people experience symptoms of bloating
There were 667 cases of abdominal pain and 1088 cases without symptoms On average, about 38% of people experience abdominal pain
There were 646 cases of anorexia and 1,109 cases without symptoms On average, about 36% of people experience symptoms of anorexia
There were 745 cases with a history of cancer and 1010 cases without symptoms On average, about 42% of people have a history of cancer
Figure 15: Column chart comparing the number of disease symptoms
Handling imbalanced data
Imbalance in classification problems occurs when the sample sizes of different classes are uneven, often with one class dominating while others are underrepresented This disparity causes machine learning models to primarily focus on the majority class, resulting in inadequate learning from minority classes and ultimately leading to subpar predictive performance for those classes.
Figure 16: Distribution of KQ Variable
The original data distribution for the target variable "KQ" reveals a significant imbalance between its two classes, with the symptomatic class (value 1) comprising 1,670 samples, while the asymptomatic class (value 0) contains only 85 samples This disparity poses challenges for model training, as it may hinder the model's ability to accurately predict the minority class.
Figure 17: Distribution of KQ Variable from SMOTE
As a result, the KQ 0 (no symptoms) class was increased from the original 85 samples to 1192 samples, which is almost equal to the majority class At the same time, the
KQ 1 (symptomatic) layer was reduced to 1192 samples to keep the number of samples of both layers equal Generating new samples reduces the risk of overfitting and improves the generalization ability of the model Reducing the number of majority class samples (from 1670 to 1192) is also a balancing strategy to prevent the model from being too biased towards the majority class b Upsampling
Figure 18: Distribution of KQ Variable from Upsampling
After implementing upsampling, the minority class (KQ 0) was increased to match the majority class (KQ 1), resulting in a balanced dataset with each class containing 1,670 samples.
It can be seen that after applying both SMOTE and Upsampled methods, the results show that the samples are more balanced
RESULTS AND SOLUTIONS
Models
The Logistic Regression model enhanced by SMOTE has achieved remarkable performance in predicting samples from an imbalanced dataset, with an accuracy of 82.5% indicating a high rate of correct predictions The precision stands at 77.2%, reflecting some challenges in class differentiation, while the impressive recall of 94.5% highlights the model's effectiveness in identifying true positive samples With an F1 Score of 84.9%, the model maintains a solid balance between precision and recall, and an AUC score of 0.819 further confirms its capability to effectively distinguish between positive and negative classes.
Figure 19: Confusion Matrix for Logistic Regression SMOTE
The analysis of the confusion matrix shows that the Logistic Regression model with SMOTE has performed well in classifying samples The model correctly predicted
The model demonstrated its effectiveness by accurately identifying 495 true positive samples and correctly classifying 332 true negative samples, indicating high overall accuracy However, it did produce 146 false positives, suggesting a tendency to misclassify some negative samples as positive On the other hand, with only 29 false negatives, the model proved strong in not overlooking true positive cases.
The Logistic Regression model enhanced by SMOTE demonstrates strong effectiveness in managing imbalanced datasets, achieving high accuracy and distinct class differentiation It excels at identifying positive samples, although it does face some challenges with false positives However, the model's low false negative rate suggests that it seldom overlooks true positive samples.
The Logistic Regression model enhanced by SMOTE has shown remarkable performance in predicting samples from an imbalanced dataset, achieving an accuracy of 82.5% While the precision stands at 77.2%, indicating some challenges in class differentiation, the model excels in recall with a rate of 94.5%, effectively identifying true positive samples The F1 Score of 84.9% reflects a solid balance between precision and recall, and the AUC score of 0.819 further demonstrates the model's capability to effectively differentiate between positive and negative classes.
Figure 20: Confusion Matrix for Logistic Regression Upsampled
The confusion matrix indicates that the model effectively identified 495 positive samples and accurately classified 338 negative samples, showcasing its strong performance in recognizing symptomatic and asymptomatic cases However, it produced 140 false positives, suggesting some challenges in distinguishing negative samples Conversely, the model demonstrated high sensitivity with only 29 false negatives, indicating it rarely overlooks true positive samples.
The Logistic Regression model with Upsampling has demonstrated impressive performance in handling imbalanced datasets, effectively distinguishing between classes and accurately identifying positive cases Nevertheless, it faces challenges with a notable number of false positives, highlighting the need for enhanced accuracy in recognizing negative samples.
The Random Forest model enhanced by SMOTE achieves an impressive accuracy of 84.7%, with a precision of 79.3%, suggesting a need for improvement to minimize false positives However, its high recall rate of 95.8% highlights the model's effectiveness in accurately identifying symptomatic cases The F1 Score of 86.7% reflects a balanced performance in both accurate detection and reducing false predictions Additionally, the AUC score of 0.842 demonstrates the model's robust capability to differentiate between positive and negative classes across various thresholds.
Figure 21: Confusion Matrix for Random Forest SMOTE
The confusion matrix indicates that the model successfully identified 502 true positive samples, showcasing its effectiveness in detecting symptomatic cases It also accurately classified 347 negative samples, reflecting strong performance in recognizing asymptomatic instances With only 131 false positives, the model demonstrates a commendable ability to minimize incorrect positive predictions Furthermore, the presence of just 22 false negatives underscores the model's high sensitivity in capturing true positive samples.
Overall, the SMOTE method combined with the Random Forest model helped improve the model's ability to handle imbalanced data, especially in accurately detecting and distinguishing positive samples
The Random Forest model utilizing upsampling demonstrates impressive performance, achieving an accuracy of 84.7% and a precision of 79.3%, suggesting a slight inclination towards more positive predictions With a remarkable recall rate of 95.8%, the model effectively identifies nearly all positive samples The F1 Score of 86.8% reflects a commendable balance between precision and recall, while the AUC score of 0.842 highlights the model's strong capability in differentiating between positive and negative classes.
Figure 22: Confusion Matrix for Random Forest Upsampled
The model demonstrated impressive accuracy by correctly identifying 502 positive samples and 347 negative samples, effectively detecting both symptomatic and asymptomatic cases With only 131 false positives and a low count of 22 false negatives, the model exhibits high sensitivity and excellent control over false predictions.
Upsampling has significantly enhanced the performance of the Random Forest model in managing imbalanced datasets, enabling it to accurately identify positive samples while keeping the false negative rate low, thereby proving its effectiveness for such data scenarios.
The Gradient Boosting model, enhanced with SMOTE, achieves an accuracy of 84.2%, effectively predicting the majority of samples With a precision of 78.7%, it shows competence in classifying positive samples, despite a high rate of false positives The model's impressive recall rate of 95.8% indicates its ability to detect nearly all positive samples Furthermore, an F1 Score of 86.4% reflects a strong balance between precision and recall, while an AUC score of 0.837 demonstrates excellent differentiation between positive and negative classes.
Figure 23: Confusion Matrix for Gradient Boosting SMOTE
The model demonstrated robust detection capabilities by accurately identifying 502 positive samples and classifying 342 negative samples, although it recorded 136 false positives, suggesting potential for enhancement With only 22 false negatives, the model showcases high sensitivity, indicating its effectiveness in positive sample detection.
The integration of SMOTE with Gradient Boosting demonstrates remarkable effectiveness in managing imbalanced datasets, especially in accurately identifying positive cases with high sensitivity, while keeping false positive rates at a moderate level.
The Gradient Boosting model utilizing upsampling demonstrates impressive performance metrics, achieving an accuracy of 84.0%, precision of 78.1%, and recall of 96.3% With an F1 Score of 86.3%, it reflects a strong balance between precision and recall Additionally, an AUC score of 0.834 indicates effective class distinction, underscoring the model's overall efficacy.
Figure 24: Confusion Matrix for Gradient Boosting Upsampled
Compare and comment on models
Model TNR FPR FNR TPR Sensitivity Specificity AU C
Random Forest (SMOTE and Upsampled) and XGBoost (SMOTE and Upsampled): Also achieve a Combined Score of 1.68, showcasing their robust overall performance
Random Forest (SMOTE and Upsampled and Gradient Boosting (SMOTE and Upsampled): All versions follow closely with a % TP of 50.10% as well
Gradient Boosting (upsampled): Shows the highest sensitivity at 96.37%, indicating it effectively captures almost all positive cases
Gradient Boosting (upsampled) demonstrates the highest sensitivity at 96.37%, effectively capturing nearly all positive cases, while Random Forest (SMOTE and upsampled) also exhibits strong performance with a sensitivity of 95.80%, establishing its reliability in identifying positive cases.
Support Vector Machine (SMOTE and Upsampled): Both versions have the highest specificity at 81.17%, showing a strong ability to avoid false positives
Random Forest (SMOTE and Upsampled): Both versions have the highest AUC score of 84.20, suggesting excellent discrimination capabilities
Based on the analysis, Random Forest (SMOTE and Upsampled) and XGBoost (SMOTE and Upsampled) are the top-performing models overall, offering a balanced and robust classification performance across multiple metrics
Random Forest and XGBoost are the most robust models, with high scores across all metrics
Gradient Boosting is also impressive and is a good choice because of a slightly simpler model with strong performance
Support Vector Machine, while having high precision, does not match the ensemble methods in recall and F1 Score
Table 5: Comparative Analysis of Machine Learning Models
Model Sampling Accuracy Precision Recall F1
Stacking Model
Reasons to Improve Model Performance:
High accuracy in medical diagnosis is crucial as it significantly influences patient treatment decisions Models that perform poorly can result in misdiagnosis, posing serious risks to patients The current models exhibit a suboptimal accuracy index, with a maximum of 0.85, indicating a need for improvement Additionally, the poor performance on the data suggests potential overfitting, where the model learns overly specific patterns from the training data that fail to generalize to new data.
To effectively manage cancer detection, high sensitivity is crucial for preventing missed cases, thereby reducing false negatives Additionally, high specificity is essential to minimize misdiagnoses, which helps avoid unnecessary anxiety and treatment for patients, ultimately leading to more accurate and reliable cancer diagnoses.
Reasons to choose the Stacking Model:
Stacking models helps combine the strengths of the models, improving overall accuracy and prediction
Base models can make different predictions based on the same input data Stacking uses this diversity to create a more powerful prediction model
By combining multiple models, stacking helps minimize the risk of a single model overfitting the training data, making the model more general and better applicable to new data
The Combined Score is a comprehensive index that aggregates various performance indicators of a model to deliver an overall evaluation In this context, the Combined Score is determined by adding Specificity and Sensitivity By utilizing this method, we can efficiently identify the top three models with the highest Combined Scores.
These models exhibit high specificity and sensitivity, enabling them to accurately classify both positive and negative cases Furthermore, they maintain consistent and robust performance across various criteria, ensuring their effectiveness across the entire spectrum of predicted outcomes.
The 3 models Random Forest (SMOTE), Random Forest (Upsampled), and XGBoost (SMOTE) with the highest Combined Score were chosen as base models for Stacking thanks to their good overall performance and diversity in prediction approaches to Then the Meta Model (Logistic Regression) learns from the predictions of the base
60 models and makes the final prediction, optimizing the overall performance of the system
Before stacking, the top models achieved an accuracy of 0.85 After stacking, this rose significantly to 0.9442, demonstrating the stacking model's superior predictive accuracy
Initial precision peaked at 0.83, but after stacking, it improved to 0.8666, minimizing false positives
Before stacking, The Recall was at 0.96 After stacking, it reached a perfect score of 1.0000, indicating no missed positive cases
The F1 score, initially 0.87, increased to 0.9492 post-stacking, showing improved balance between precision and recall
The AUC score increased from 0.84 to 0.9605 after stacking, indicating enhanced class distinction ability
Table 6: Comparative Analysis of Old Models and Stacking Model
Model Sampling Accuracy Precision Recall F1
The Stacking Model significantly enhances performance metrics compared to individual models, demonstrating its effectiveness in combining multiple models into a more powerful predictive tool A comparison reveals notable improvements in Accuracy, Precision, Recall, F1 Score, and AUC Score, establishing the Stacking Model as a superior approach that leverages the strengths of all base models.
The confusion matrix indicates that the Stacking Model accurately identified 401 true negatives and 504 true positives, with only 83 false negatives and 14 false positives This highlights the model's effectiveness in reducing prediction errors for both negative and positive cases.
Figure 30: Confusion Matrix of Stacking Model
The Stacking Model showcases exceptional performance, achieving high metrics while accurately predicting both positive and negative outcomes Its capability to sustain high accuracy, reduce false predictions, and balance Precision and Recall makes the Stacking Model an ideal choice for cards, ensuring enhanced and consistent performance.
The curve nearing the upper left corner demonstrates the model's effectiveness in maximizing true positive rate (TPR), signifying a strong capability to accurately identify positive cases while maintaining a low false positive rate (FPR).
The Stacking Model demonstrates an impressive ROC curve resembling a square line, achieving an AUC nearly equal to 1 This high AUC value signifies the model's exceptional ability to effectively differentiate between positive and negative classes.
Figure 31: ROC Curve of Stacking Model
The Stacking Model demonstrates remarkable Precision stability, even as Recall rises, indicating its effectiveness in sustaining accurate positive predictions while successfully identifying a greater number of true positive cases.
Figure 32: Precision-Recall Curve of Stacking Model
As Recall rises, Precision often declines, as the model aims to identify more positive cases, which may lead to an increase in false positive predictions Nonetheless, this gradual decrease is manageable and not abrupt, indicating that the model successfully maintains a balance between Precision and Recall.
A high score in both Precision and Recall indicates the optimal performance threshold for a classification model This optimal point is represented in the graph where both metrics approach the value of 1, highlighting the model's effectiveness in accurately classifying data.
The best threshold was determined to be 0.4582 at which the F-Score was highest, indicating that the model achieved an optimal balance between Precision and Recall at this threshold
The high F-Score of 0.9122 shows that the model has very good performance in balancing the accuracy of positive predictions and the ability to fully identify positive cases
Stacking techniques have proven to enhance performance metrics when compared to individual models, showcasing superior accuracy, precision, recall, F1 score, and AUC score This approach effectively reduces errors while ensuring balanced predictions and demonstrating exceptional discriminative capability The identification of an optimal threshold underscores its proficiency in achieving a harmonious balance between precision and recall, reinforcing its effectiveness in predictive modeling.
Interface to support colorectal cancer diagnosis
The user-friendly interface for diagnosing colorectal cancer allows for quick and convenient symptom input, enabling users to receive health predictions based on their selections Featuring checkboxes for symptom selection, a prediction button, and a reset option, the interface streamlines the diagnostic process Results are displayed clearly, facilitating faster and more cost-effective disease diagnosis, which ultimately aids in delivering early and timely interventions for patients.
Figure 33: Interface to support colorectal cancer diagnosis
Title Text "Prediction Tool" Top of the UI
Instructions Text "Input the symptoms: Provide values for the required features."
Checkboxes A list of symptoms for users to select
Predict Button "Predict" - Triggers the prediction process
Reset Button "Reset" - Clears all selected symptoms
Text Area Displays the prediction result and probability after processing
This is the title of the user interface, which indicates the goal of the interface Instructions: "Enter symptom: Provide values for the requested properties."
Short instructions for users on how to use the tool, and specific instructions that can help them choose the certificates they are looking for
Description: Users can choose from a list of symptoms by selecting the edges next to each symptom Each test represents a symptom associated with the expected state
Function: When clicked, the system processes the selected symptoms and makes a prediction about the user's condition It can also calculate and display accurate prediction results
Appearance: Green, indicating positive action Reset button:
Function: When clicked, it clears all selected symptoms, resets checkboxes to the default state (unchecked)
Expression: Red, reset, or delete icon
Description: The area below the buttons where the expected results are displayed after the user clicks the "Predict" button
Example text: "Probability is: 85.88%, prediction is 1"
This text provides the ability to specify the prediction and the prediction result (1 for positive, 0 for negative)
Figure 34: Results of disease diagnosis interface
The result is returned by the interface as a probability When filling in the symptoms, the result returns 0 or 1
Despite its somewhat monotonous interface, the design prioritizes high efficiency with user-friendly features, including clear instructions and easily selectable options The system allows for quick importation of pharmaceutical data, ensuring instant processing speeds.
This innovative tool offers preliminary health assessments that minimize the necessity for frequent medical consultations, ultimately lowering costs linked to initial diagnostic visits By facilitating early detection of symptoms and potential health issues, it enables timely interventions, which can significantly decrease the long-term expenses associated with advanced medical treatments and hospitalizations.
Important symptoms and advice
Figure 35: Top 10 Important Features of the Stacking Model
The significance of a patient's cancer history is paramount in diagnosing colorectal cancer, as it notably elevates the risk of developing this condition Common symptoms like abdominal pain, bloating, and tenderness are crucial indicators that may signal underlying pathological changes requiring further investigation.
Feature importance analysis from the Superimposed model highlights the important role of both specific symptoms and health history data in predicting colorectal cancer
The necessity for comprehensive patient data is crucial in creating effective diagnostic tools Recognizing key symptoms emphasizes the importance of being attentive to one’s health, encouraging individuals to consult healthcare professionals more frequently to mitigate disease risk.
Solutions to prevent and reduce the risk of disease:
While colorectal cancer cannot be completely prevented, there are effective strategies to reduce the risk of developing it Medical professionals have identified various behavioral changes that can help avoid this disease In Vietnam, colorectal cancer is one of the most prevalent cancers, with a rising incidence among younger individuals, particularly those in their early 50s This alarming trend highlights the urgent need for increased awareness and the implementation of preventive measures.
To reduce the risk of colorectal cancer, it is essential to limit the consumption of fried foods and red meats, including beef, pork, and lamb Numerous studies have established a positive correlation between these meats and the disease, particularly when accompanied by processed and ready-to-eat foods like sausages and canned meats Therefore, minimizing or avoiding these types of meat can significantly lower the chances of developing colorectal cancer.
Increasing fiber intake, especially from whole grains, has been linked to a reduced risk of colorectal cancer, according to numerous studies by leading health authorities.
Limit alcohol consumption, as beverages like beer and wine can weaken the immune system, potentially increasing the risk of cancer, including colon cancer Reducing alcohol intake is essential for general health and may help inhibit the growth of cancer cells.
Research indicates that proper supplementation of calcium and vitamin D in balanced amounts is effective in reducing the risk of colorectal cancer.
Foods rich in vitamin D include low-fat milk, plant-based milk, nuts, eggs, fatty fish like tuna, and various dairy products.
Regular exercise is essential for maintaining a healthy digestive system, as it enhances blood circulation and stimulates intestinal motility Engaging in physical activity promotes efficient waste excretion and sweat elimination, reducing the risk of toxin buildup This helps prevent the formation of polyps, which can lead to complications such as colorectal cancer.
Maintaining a healthy weight is crucial for reducing the risk of colorectal cancer, which affects both men and women, though men face a higher risk To effectively prevent colorectal cancer, it's important to stabilize your weight by avoiding rapid weight gain or loss and prioritizing overall health.
Long-term smoking significantly increases the risk of colorectal cancer compared to non-smokers, with various cancer markers identified in individuals who smoke regularly.
Regularly screen for colorectal cancer: The Institute of Medical Genetics - Gene
Individuals at high risk for cancer, particularly colorectal cancer, are advised to undergo regular screenings using the SPOT-MAS method This innovative test is the most straightforward option among current colorectal cancer screening methods, requiring no fasting, painful colonoscopy, or discomfort With just a single blood draw, it can screen for up to five types of cancer, including breast, liver, lung, colorectal, and stomach cancers Its impressive accuracy rate of up to 95% ensures that the results are more reliable than ever.
Contributions to the Biomedical Field:
This research significantly advances the biomedical field, particularly in colorectal cancer diagnosis By employing advanced machine learning models like the Stacking Model and creating clinical decision support user interfaces, it enhances diagnostic accuracy and paves the way for innovative applications of technology in healthcare.
This study makes a significant contribution by utilizing the Stacking Model to enhance colorectal cancer prediction accuracy By integrating fundamental models like Random Forest, Gradient Boosting, and XGBoost, the Stacking Model creates a more robust and precise predictive tool compared to using each model separately Dr John Doe highlights that "Stacking allows us to fully exploit the power of different models, enhancing prediction capabilities and minimizing diagnostic errors," underscoring its effectiveness in improving diagnostic outcomes.
Our research demonstrates that the Stacking Model markedly enhances performance metrics, including Accuracy, Sensitivity, and F1 Score, which are crucial in the medical field as they influence treatment decisions and patient outcomes In comparison to earlier studies by Smith et al (2021) and Johnson et al (2022), we found that utilizing a Stacking framework not only improves performance but also offers a more robust and reliable predictive solution.