An effective multi-step feature selection framework for clinical outcome prediction using electronic medical records.

Abstract

Background: Identifying key variables is essential for developing clinical outcome prediction models based on high-dimensional electronic medical records (EMR). However, despite the abundance of feature selection (FS) methods available, challenges remain in choosing the most appropriate method, deciding how many top-ranked variables to include, and ensuring these selections are meaningful from a medical perspective.

Methods: We developed a practical multi-step feature selection (FS) framework that integrates data-driven statistical inference with a knowledge verification strategy. This framework was validated using two distinct EMR datasets targeting different clinical outcomes. The first cohort, sourced from the Medical Information Mart for Intensive Care III (MIMIC-III), focused on predicting acute kidney injury (AKI) in ICU patients. The second cohort, drawn from the MIMIC-IV Emergency Department (MIMIC-IV-ED), aimed to estimate in-hospital mortality (IHM) for patients transferred from the ED to the ICU. We employed various machine learning (ML) methods and conducted a comparative analysis considering accuracy, stability, similarity, and interpretability. The effectiveness of our FS framework was evaluated using discrimination and calibration metrics, with SHAP applied to enhance the interpretability of model decisions.

Results: Cohort 1 comprised 48,780 ICU encounters, of which 8,883 (18.21%) developed AKI. Cohort 2 included 29,197 transfers from the ED to the ICU, with 3,219 (11.03%) resulting in IHM. Among the ten ML methods evaluated, the tree-based ensemble method achieved the highest accuracy. As the number of top-ranking features increased, the models' accuracy began to stabilize, while feature subset stability (considering sample variations) and inter-method feature similarity reached optimal levels, confirming the validity of the FS framework. The integration of interpretative methods and expert knowledge in the final step further improved feature interpretability. The FS framework effectively reduced the number of features (e.g., from 380 to 35 for Cohort 1, and from 273 to 54 for Cohort 2) without significantly affecting prediction performance (Delong test, p > 0.05).

Conclusion: The multi-step FS method developed in this study successfully reduces the dimensionality of features in EMR while preserving the accuracy of clinical outcome prediction. Furthermore, it improves the interpretability of risk factors by incorporating expert knowledge validation.