Abstract
It is important for physicians' clinical decision support to classify coronary heart disease (CHD). Customizing personalized predictive models for patients requires selecting a patient group from an existing medical database that most closely resembles the indexed patients. In this study, we introduce a new concept using patient similarity for the classification of patients with CHD. We performed a structured representation of CHD patients and obtained a multidimensional attribute distance matrix between patient pairs. Similarity between patient pairs was predicted using machine learning (ML) models to determine clinical outcomes for indexed patients based on matched similar patients. The new measure shows marked improvements over traditional classification measures. LightGBM emerged as the top-performing ML model, achieving 88.52% accuracy. The medical applications of ML supported by similarity analytics represent a promising solution to reduce physician workload and achieve the goal of “precision medicine.”
Introduction
Coronary Heart Disease (CHD) remains a leading global health concern characterized by complex and varied clinical manifestations. While machine learning (ML) has been widely applied to CHD prediction, traditional models often focus on general population trends, making them difficult to tailor for individual patients.
Standard statistical models often overlook the fundamental way physicians operate: through analogical reasoning. When evaluating a new patient, clinicians instinctively compare the patient’s symptoms, history, and lab results to previous cases. Current clinical decision support systems often lack the ability to quantitatively measure the "distance" between two patients to provide truly personalized "precision medicine." This research addresses this gap by proposing a personalized approach to classifying CHD subtypes through patient similarity analysis.
Methodology
The researchers developed a comprehensive pipeline to transform raw Electronic Medical Records (EMRs) into a predictive similarity engine.
Data Processing and Representation
The study utilized a dataset of 4,910 EMRs from a tertiary hospital in Xinjiang, China. These records were manually annotated and classified into four CHD subtypes: Stable angina pectoris, Unstable angina pectoris, Ischemic cardiomyopathy, and Myocardial infarction.
- Structured Representation: 37 key attributes were identified from EMRs, encompassing demographics, symptoms (e.g., chest pain quality), laboratory tests (e.g., Troponin T, BNP), risk factors, and imaging data.
- CatBoost Encoding: To manage the heterogeneity of medical data, the CatBoost-Encoder was employed. This method is specifically designed to handle categorical features with high efficiency, significantly outperforming standard one-hot encoding.
- Distance Calculation: A multidimensional attribute distance matrix was calculated for patient pairs. Labels were assigned based on a binary "consistency" metric—1 if the diagnosis matched, and 0 otherwise.
Model Selection
The core prediction engine utilizes LightGBM (Light Gradient Boosting Machine). This model was selected for its superior ability to handle large-scale data with faster training speeds and lower memory consumption compared to traditional Gradient Boosting Decision Trees (GBDT).
Results and Discussion
The patient-to-patient similarity approach aligns closely with clinical workflows, shifting the focus from general classification to finding the most relevant historical precedents.
Performance Metrics
The LightGBM model demonstrated superior performance across all evaluated metrics:
- Accuracy: 88.52%
- Precision: 92.57%
- Recall: 93.01%
- F1-score: 92.79%
The custom distance calculation method (88.52% accuracy) significantly outperformed traditional Euclidean (81.77%) and Jaccard (80.20%) distance metrics, indicating that simple geometric distances are insufficient for capturing the nuances of patient similarity.
Computational Efficiency
In clinical environments, processing speed is critical. LightGBM achieved a runtime of only 15.9 seconds, making it substantially faster than competitors like K-Nearest Neighbors (37.82s) and Random Forest (43.56s). This efficiency makes it a viable candidate for real-time clinical decision support.
Conclusion
This study demonstrates that patient similarity analysis, powered by LightGBM and sophisticated data encoding, provides a robust framework for CHD classification. By mimicking the analogical reasoning of physicians, the model moves closer to the ideal of precision medicine.
While currently limited to CHD subtypes and data from a single institution, future work will focus on expanding the dataset to improve generalizability and exploring applications in other medical domains. The integration of ML-supported similarity analytics represents a promising path toward reducing physician workload while enhancing diagnostic accuracy.
Summary updated