[Paper Review] MedLens: Improve Mortality Prediction Via Medical Signs Selecting and Regression

Xuesong Ye, Jun Wu+2 more|arXiv (Cornell University)|May 19, 2023

Machine Learning in Healthcare26 references4 citations

TL;DR

MedLens improves ICU mortality prediction by selecting high-frequency, high-correlation medical signs and applying a Random Forest-based interpolation method to handle missing time-series data, achieving 0.9553 AUC-ROC and 0.8051 AUC-PR—significantly outperforming prior benchmarks.

ABSTRACT

Monitoring the health status of patients and predicting mortality in advance is vital for providing patients with timely care and treatment. Massive medical signs in electronic health records (EHR) are fitted into advanced machine learning models to make predictions. However, the data-quality problem of original clinical signs is less discussed in the literature. Based on an in-depth measurement of the missing rate and correlation score across various medical signs and a large amount of patient hospital admission records, we discovered the comprehensive missing rate is extremely high, and a large number of useless signs could hurt the performance of prediction models. Then we concluded that only improving data-quality could improve the baseline accuracy of different prediction algorithms. We designed MEDLENS, with an automatic vital medical signs selection approach via statistics and a flexible interpolation approach for high missing rate time series. After augmenting the data-quality of original medical signs, MEDLENS applies ensemble classifiers to boost the accuracy and reduce the computation overhead at the same time. It achieves a very high accuracy performance of 0.96 AUC-ROC and 0.81 AUC-PR, which exceeds the previous benchmark.

Motivation & Objective

• To address the critical issue of poor data quality in electronic health records (EHR), particularly high missing rates and low-correlation medical signs, which hinder mortality prediction performance.
• To identify and retain only the most informative medical signs—those with high recording frequency and strong correlation to mortality—thereby reducing model complexity and noise.
• To develop a novel, flexible interpolation method based on Random Forest regression that effectively reconstructs missing values in irregularly sampled time series.
• To improve prediction accuracy and computational efficiency by combining high-quality data preprocessing with lightweight ensemble classification.

Proposed method

• Conducted a comprehensive analysis of the MIMIC-III dataset to measure missing rates and correlation scores across 100+ medical signs, identifying low-frequency and weakly correlated signals as noise.
• Implemented a statistical feature selection method that filters out long-tail medical signs with low recording frequency and weak correlation to mortality, retaining only high-impact signals.
• Designed a Random Forest regression-based interpolation technique that learns temporal patterns and fills missing values in multivariate time series, outperforming baseline forward/backward filling.
• Applied an ensemble classifier (Random Forest and Gradient Boosting) to the cleaned and interpolated time series data to enhance prediction accuracy and reduce computational overhead.
• Used AUC-ROC and AUC-PR as primary evaluation metrics to benchmark performance against state-of-the-art models.
• Optimized for efficiency by selecting Random Forest as the final classifier due to its balance of speed and accuracy, achieving 97.72% of the accuracy of Gradient Boosting in just 2.28% of the time.

Experimental results

Research questions

• RQ1 How do missing rates and correlation scores of medical signs in EHR data affect the performance of mortality prediction models?
• RQ2 Can a statistical feature selection method improve prediction accuracy by filtering out low-frequency and weakly correlated medical signs?
• RQ3 Does a Random Forest-based interpolation method for time series with high missing rates lead to better prediction performance than traditional interpolation techniques?
• RQ4 Can a lightweight ensemble classifier maintain high accuracy while reducing computational cost in mortality prediction tasks?

Key findings

• The comprehensive missing rate across medical signs in the MIMIC-III dataset is extremely high, with many signs being rarely recorded and largely irrelevant to mortality.
• Only a small subset of high-frequency, high-correlation medical signs is sufficient to achieve very high prediction accuracy, demonstrating that data quality is more critical than model complexity.
• The proposed Random Forest interpolation method significantly outperforms baseline forward/backward filling, improving AUC-ROC by 8.5% and AUC-PR by 47% compared to Harutyunyan et al. [1].
• MedLens achieved 0.9553 AUC-ROC and 0.8051 AUC-PR, surpassing the benchmark model by Harutyunyan et al. [1], which reported only 0.87 AUC-ROC and 0.533 AUC-PR.
• Random Forest as the final classifier achieved 97.72% of the accuracy of Gradient Boosting while requiring only 2.28% of the computation time, making MedLens highly efficient for large-scale deployment.
• The study confirms that improving data quality—through feature selection and advanced interpolation—can lead to substantial performance gains even without modifying the underlying machine learning algorithm.