[Paper Review] Using Latent Class Analysis to Identify ARDS Sub-phenotypes for Enhanced Machine Learning Predictive Performance

Tony Wang, Tim Tschampel+2 more|arXiv (Cornell University)|Mar 28, 2019

Machine Learning in Healthcare19 references4 citations

TL;DR

This study uses latent class analysis (LCA) to identify three distinct ARDS sub-phenotypes from MIMIC-III ICU data, enabling improved machine learning prediction of mortality. By training separate models on each sub-phenotype, AUCs reached 0.986 for the highest-mortality group, significantly outperforming models trained on the entire heterogeneous ARDS population.

ABSTRACT

In this work, we utilize Machine Learning for early recognition of patients at high risk of acute respiratory distress syndrome (ARDS), which is critical for successful prevention strategies for this devastating syndrome. The difficulty in early ARDS recognition stems from its complex and heterogenous nature. In this study, we integrate knowledge of the heterogeneity of ARDS patients into predictive model building. Using MIMIC-III data, we first apply latent class analysis (LCA) to identify homogeneous sub-groups in the ARDS population, and then build predictive models on the partitioned data. The results indicate that significantly improved performances of prediction can be obtained for two of the three identified sub-phenotypes of ARDS. Experiments suggests that identifying sub-phenotypes is beneficial for building predictive model for ARDS.

Motivation & Objective

• Address the challenge of poor predictive performance in ARDS due to clinical heterogeneity.
• Investigate whether partitioning ARDS patients into biologically coherent sub-phenotypes improves machine learning model performance.
• Utilize latent class analysis (LCA) to identify homogeneous subgroups based on clinical variables such as labs, vitals, and comorbidities.
• Compare predictive performance of models trained on sub-phenotypes versus the entire ARDS cohort.
• Evaluate whether sub-phenotype-specific models yield higher discrimination (AUC) and better calibration for mortality prediction.

Proposed method

• Apply latent class analysis (LCA) to MIMIC-III ICU data to identify three distinct ARDS sub-phenotypes based on clinical variables including PaO2/FiO2, PEEP, lactate, bilirubin, creatinine, WBC, and comorbidities.
• Use Berlin criteria to define ARDS onset and extract clinical features (e.g., mean bicarbonate, plateau pressure, minimum SBP, maximum heart rate) within 24 hours post-onset.
• Train gradient boosting machines (GBM) and random forest models separately on each LCA-identified sub-phenotype to assess predictive performance.
• Evaluate model performance using AUC, sensitivity, specificity, positive and negative predictive values, and accuracy, with bootstrapped confidence intervals.
• Compare sub-phenotype model AUCs to the reference model trained on the entire ARDS population.
• Use statistical testing (p-values) to assess significance of AUC differences between sub-phenotype models and the overall model.

Experimental results

Research questions

• RQ1 Can latent class analysis effectively identify biologically meaningful sub-phenotypes within the heterogeneous ARDS population?
• RQ2 Does training machine learning models on LCA-identified sub-phenotypes improve predictive performance for ARDS mortality compared to modeling the entire ARDS cohort?
• RQ3 Which ARDS sub-phenotype exhibits the highest mortality and most severe clinical markers, and can it be reliably predicted?
• RQ4 What is the relative improvement in AUC and other performance metrics when using sub-phenotype-specific models versus a unified model?
• RQ5 How does the distribution of key clinical variables (e.g., bilirubin, creatinine, PaO2/FiO2) differ across the identified sub-phenotypes?

Key findings

• Latent class analysis identified three distinct ARDS sub-phenotypes with differing clinical characteristics and mortality rates, with phenotype 2 showing the highest mortality (48%).
• Sub-phenotype 2, characterized by high bilirubin, creatinine, thrombocytopenia, and low systolic blood pressure, had an AUC of 0.986 using GBM, significantly outperforming the overall model (AUC 0.904).
• Sub-phenotype 1 achieved an AUC of 0.983 with GBM, also significantly better than the overall model (p < 0.0001).
• Sub-phenotype 3 showed only modest improvement, with an AUC of 0.909, not significantly different from the overall model (p = 0.6538).
• The sub-phenotype-specific models achieved near-perfect NPV (1.000) and high sensitivity (1.000 for phenotype 1), indicating strong ability to rule out non-events.
• The study demonstrates that accounting for ARDS heterogeneity via LCA leads to substantial performance gains in machine learning prediction, particularly for high-mortality subgroups.