Abstract
Predicting stroke risk is a critical task in clinical medicine, traditionally handled by the Cox proportional hazards model. However, high-dimensional medical data with significant missingness pose challenges for traditional statistical methods. This paper presents an integrated machine learning pipeline featuring a "conservative mean" (CM) feature selection heuristic and a novel margin-based censored regression (MCR) algorithm. Evaluated on the Cardiovascular Health Study (CHS) dataset, the proposed approach achieves an AUC of 0.777, representing a 16% reduction in prediction error over standard models, while identifying novel clinical risk factors such as cognitive and lung function measures.
Introduction
Stroke remains a leading cause of mortality and long-term disability worldwide. For decades, the medical community has relied on the Cox proportional hazards model for risk assessment, typically utilizing a small set of expert-selected features like age and blood pressure. However, modern medical datasets are high-dimensional, often containing thousands of features and significant amounts of missing data. Traditional models struggle with this scale and often require manual feature selection, which is slow and potentially biased.
The challenge lies in building a system that can automatically handle missing data, select the most robust features from noisy datasets, and predict not just the occurrence of a stroke, but its relative risk and timing—a task complicated by "censored" data where events have not yet happened for some subjects at the time of study.
Methodology
The researchers proposed a three-stage pipeline designed to handle the complexities of clinical data.
Data Imputation
The study compared several methods to fill in missing values, including column mean/median, linear regression, and Expectation Maximization (EM). Surprisingly, the authors found that column median imputation led to the best overall prediction quality, despite linear regression having higher raw imputation accuracy. This suggests that for downstream machine learning tasks, simple and robust imputation may be preferable to complex models that might introduce their own biases.
Conservative Mean (CM) Feature Selection
A key contribution of this work is the Conservative Mean heuristic. Instead of selecting features based on their average performance across data folds, the authors used the heuristic $(\mu - \sigma)$, where $\mu$ is the mean AUC and $\sigma$ is the standard deviation. This approach penalizes features with high variance, ensuring that the selected features are robust and generalize well across different subsets of patients.
Margin-based Censored Regression (MCR)
While Support Vector Machines (SVMs) are effective for binary classification, they often ignore the timing of events. MCR unifies linear regression with an SVM-like margin. It attempts to keep predicted "hazard values" close to the actual time-to-stroke while maintaining a clear margin between those who had a stroke and those who did not, effectively handling censored medical data.
Experimental Results
The study utilized the Cardiovascular Health Study (CHS) dataset, a benchmark in medical research involving approximately 5,000 individuals over age 65. After preprocessing, the model utilized 796 features to predict 5-year stroke risk.
Predictive Performance
The integrated MCR + CM approach significantly outperformed the state-of-the-art. The model achieved an AUC of 0.777, compared to 0.734 for the standard Cox model. This represents a 16% reduction in prediction error. Furthermore, in terms of ranking relative risk (Concordance Index), the MCR model achieved a score of 0.770, outperforming the Cox model’s 0.730.
Efficiency and Discovery
The CM feature selection was highly efficient, taking less than 10 minutes to process 796 features, whereas traditional forward selection took roughly 60 hours on the same hardware. Beyond speed, the automated approach served as a discovery tool, identifying risk factors not found in traditional 16-factor clinical models, such as "Number of symbols correctly coded" (a cognitive measure) and "FVC percent predicted" (a lung function measure).
Conclusion
This research demonstrates that an integrated machine learning approach can significantly improve the accuracy of stroke prediction while uncovering non-obvious clinical risk factors. The Conservative Mean heuristic provides a robust method for feature selection in non-homogeneous medical data. Future work may involve hybrid approaches using $L_1$ regularization to prune redundant features before applying the CM heuristic, further refining the selection of the most robust predictors for clinical use.
Summary updated