[Paper Review] PPMF: A Patient-based Predictive Modeling Framework for Early ICU Mortality Prediction

Mohammad Amin Morid, Olivia R. Liu Sheng+1 more|arXiv (Cornell University)|Apr 24, 2017

Machine Learning in Healthcare4 citations

TL;DR

PPMF is a patient-centric machine learning framework that enhances early ICU mortality prediction by modeling dynamic patient trajectories using 48-hour time-series data. It integrates temporal feature learning, local similarity-based classification, and gradient-based feature weighting, achieving superior performance over traditional scores and aggregation methods on MIMIC-III data.

ABSTRACT

To date, developing a good model for early intensive care unit (ICU) mortality prediction is still challenging. This paper presents a patient based predictive modeling framework (PPMF) to improve the performance of ICU mortality prediction using data collected during the first 48 hours of ICU admission. PPMF consists of three main components verifying three related research hypotheses. The first component captures dynamic changes of patients status in the ICU using their time series data (e.g., vital signs and laboratory tests). The second component is a local approximation algorithm that classifies patients based on their similarities. The third component is a Gradient Decent wrapper that updates feature weights according to the classification feedback. Experiments using data from MIMICIII show that PPMF significantly outperforms: (1) the severity score systems, namely SASP III, APACHE IV, and MPM0III, (2) the aggregation based classifiers that utilize summarized time series, and (3) baseline feature selection methods.

Motivation & Objective

• To address the challenge of early ICU mortality prediction using longitudinal clinical data.
• To improve upon traditional severity scores (e.g., SAPS III, APACHE IV) by leveraging dynamic patient trajectories.
• To develop a framework that captures patient-specific temporal patterns and adapts feature importance through feedback.
• To outperform aggregation-based classifiers and baseline feature selection methods in predictive accuracy.
• To enable personalized, real-time risk assessment within the first 48 hours of ICU admission.

Proposed method

• The framework processes time-series data (vital signs, lab tests) from the first 48 hours post-ICU admission to model patient status evolution.
• It employs a local approximation algorithm to classify patients based on similarity in their clinical trajectory patterns.
• A gradient descent wrapper dynamically updates feature weights using classification feedback to optimize predictive performance.
• The model integrates temporal dynamics, local patient clustering, and adaptive feature weighting into a unified pipeline.
• It uses MIMIC-III as the primary dataset for training and evaluation, with standardized preprocessing and feature engineering.
• Performance is evaluated using standard metrics such as AUC-ROC, comparing against severity scores and baseline models.

Experimental results

Research questions

• RQ1 Can a patient-based modeling framework improve early ICU mortality prediction compared to traditional severity scores?
• RQ2 Does incorporating dynamic time-series data enhance predictive performance over static or aggregated summaries?
• RQ3 Can local similarity-based classification combined with adaptive feature weighting outperform baseline feature selection methods?
• RQ4 How does the integration of temporal modeling and feedback-driven feature weighting affect model robustness and accuracy?
• RQ5 To what extent does PPMF generalize across diverse ICU patient populations in real-world data?

Key findings

• PPMF significantly outperforms SAPS III, APACHE IV, and MPM0III in predicting ICU mortality, with higher AUC-ROC values.
• The framework achieves better performance than aggregation-based classifiers that summarize time-series data into static features.
• Baseline feature selection methods are outperformed by PPMF’s gradient-based feature weighting mechanism.
• The integration of temporal dynamics and local similarity modeling leads to improved discrimination of high-risk patients.
• The model demonstrates robustness and generalization on real-world MIMIC-III ICU data, validating its clinical relevance.
• The proposed framework provides a scalable, patient-centric approach to early mortality prediction with actionable clinical insights.