[Paper Review] AdaCare: Explainable Clinical Health Status Representation Learning via Scale-Adaptive Feature Extraction and Recalibration

Liantao Ma, Junyi Gao+7 more|arXiv (Cornell University)|Nov 27, 2019

Machine Learning in Healthcare27 references4 citations

TL;DR

AdaCare is a deep learning model for clinical health status representation learning that uses scale-adaptive dilated convolutions and feature recalibration to capture multi-scale biomarker trends and adaptively highlight high-risk features. It achieves state-of-the-art prediction performance on ESRD and MIMIC-III datasets while providing clinically interpretable, expert-verified insights.

ABSTRACT

Deep learning-based health status representation learning and clinical prediction have raised much research interest in recent years. Existing models have shown superior performance, but there are still several major issues that have not been fully taken into consideration. First, the historical variation pattern of the biomarker in diverse time scales plays a vital role in indicating the health status, but it has not been explicitly extracted by existing works. Second, key factors that strongly indicate the health risk are different among patients. It is still challenging to adaptively make use of the features for patients in diverse conditions. Third, using prediction models as the black box will limit the reliability in clinical practice. However, none of the existing works can provide satisfying interpretability and meanwhile achieve high prediction performance. In this work, we develop a general health status representation learning model, named AdaCare. It can capture the long and short-term variations of biomarkers as clinical features to depict the health status in multiple time scales. It also models the correlation between clinical features to enhance the ones which strongly indicate the health status and thus can maintain a state-of-the-art performance in terms of prediction accuracy while providing qualitative interpretability. We conduct a health risk prediction experiment on two real-world datasets. Experiment results indicate that AdaCare outperforms state-of-the-art approaches and provides effective interpretability, which is verifiable by clinical experts.

Motivation & Objective

• Address the lack of multi-scale temporal pattern modeling in existing clinical representation learning methods.
• Enable adaptive feature weighting based on patient-specific health trajectories and risk profiles.
• Provide high-fidelity, clinically verifiable interpretability to enhance trust in black-box prediction models.
• Improve mortality risk prediction accuracy by integrating long-term trends and short-term anomalies in biomarkers.

Proposed method

• Employ multi-scale dilated convolutions to extract long-term and short-term variation patterns of clinical biomarkers across different time scales.
• Introduce a scale-adaptive feature recalibration module inspired by the squeeze-and-excitation (SE) block to dynamically reweight features based on their predictive importance.
• Use learnable attention mechanisms with sigmoid and sparsemax activation functions to emphasize high-risk features in the latest visit.
• Integrate temporal feature representations into a unified health status embedding for downstream prediction tasks.
• Train the model end-to-end using cross-entropy loss on mortality and disease progression prediction tasks.
• Validate interpretability through feature importance analysis and expert evaluation by nephrologists.

Experimental results

Research questions

• RQ1 Can multi-scale temporal feature extraction improve the representation of clinical health status in longitudinal EHR data?
• RQ2 How does adaptive feature recalibration enhance prediction performance across diverse patient conditions?
• RQ3 To what extent can the model’s attention mechanism highlight clinically meaningful biomarkers that align with expert knowledge?
• RQ4 Does the integration of scale-adaptive feature recalibration lead to both higher accuracy and more interpretable predictions compared to SOTA baselines?

Key findings

• AdaCare outperforms state-of-the-art models on both the ESRD and MIMIC-III datasets in mortality risk prediction, with statistically significant improvements.
• The variant with multi-scale dilated convolutions and sigmoid-based recalibration (AdaCare$_{c, ho}$) achieves the highest AUC, confirming the benefit of multi-scale temporal modeling.
• The sparsemax-based recalibration variant (AdaCare$_{c, ho}$) enhances only the most predictive features, maintaining strong performance while improving interpretability by suppressing irrelevant features.
• Clinical experts confirmed that the model’s attention weights align with established medical knowledge, particularly for serum albumin, urea, and serum chlorine in ESRD patients.
• The model learned novel, previously underexplored associations—such as serum chlorine’s link to cachexia and infection—validated by nephrologists as clinically relevant.
• Quantitative analysis of feature importance showed consistent, biologically plausible patterns across different causes of death, supporting the model’s reliability and clinical relevance.