Source-linked AI summary
AdaCare: Explainable Clinical Health Status Representation Learning via Scale-Adaptive Feature Extraction and Recalibration
Liantao Ma, Junyi Gao, Yasha Wang, Chaohe Zhang, Jiangtao Wang, Wenjie Ruan, Wen Tang, Xin Gao, Xinyu Ma
TL;DR
Existing clinical representation models do not fully capture biomarker variation across time scales, adapt feature use to diverse patients, or provide interpretable predictions alongside strong performance. AdaCare combines multi-scale biomarker extraction with adaptive feature recalibration, and outperforms baselines on two prediction tasks while providing clinically confirmed interpretability.
Problem
Existing models do not fully capture multi-scale biomarker variation, adapt feature use across patient conditions, or provide satisfying interpretability with high prediction performance.
Method
AdaCare uses multi-scale dilated convolution and scale-adaptive feature recalibration to represent health status and identify the most indicative clinical features.
Results
AdaCare outperforms baseline approaches on decompensation and mortality prediction across the MIMIC-III and ESRD datasets, while its medical knowledge is positively confirmed by clinical experts.
Takeaways & Limitations
AdaCare provides feature-level interpretability alongside prediction performance for health-risk assessment across diverse clinical conditions.
Abstract
from arXiv · showhide
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.
Introduction
Clinical health-status prediction must represent changing biomarker patterns across time scales, adapt feature use to patient conditions, and remain interpretable. AdaCare addresses these issues with multi-scale extraction and adaptive recalibration, while outperforming baselines in two real-world prediction tasks.
- Health-status prediction from longitudinal EMR can support individualized treatment, adverse-outcome prevention, resource allocation, and reduced medical cost.
- Existing methods do not comprehensively embed temporal health information or provide verifiable interpretations.
- Biomarker variation must be captured at multiple time scales because long-term trends and short-term abnormalities indicate different aspects of health risk.
- Risk-indicating features differ across patients, including by disease condition and biomarker behavior, requiring adaptive feature use.
- AdaCare uses multi-scale dilated convolution to capture long- and short-term biomarker patterns and build more comprehensive health-status representations.
- Its scale-adaptive recalibration models feature relationships to enhance high-risk features, suppress less useful ones, and provide fine-grained feature-level interpretation.
- AdaCare outperforms baseline approaches on decompensation and mortality prediction across the MIMIC-III and ESRD datasets, with feature knowledge positively confirmed by clinical experts.
Related Work
Prior clinical prediction models use EMR representation learning and attention or convolution mechanisms, but important gaps remain. Existing work has not jointly achieved effective multi-scale biomarker modeling, dynamic biomarker importance, and improved predictive performance in an end-to-end framework.
- Deep learning models commonly embed EMR into low-dimensional patient health-status representations for mortality, diagnosis, and phenotype prediction.
- Convolution-based approaches extract high-level temporal clinical features, but prior work does not effectively evaluate biomarker variation across multiple time scales.
- Attention-based models mainly provide visit-level interpretability by measuring relationships among clinical visits.
- Medical-feature-level methods identify disease-code or feature importance, but some are not end-to-end and RETAIN has limited predictive performance relative to the basic GRU model.
- Existing studies still do not dynamically capture biomarker importance while also improving performance in an end-to-end healthcare prediction model.
Preliminary
The paper motivates AdaCare with dynamic health-risk prediction for ESRD patients receiving periodic clinical testing. It formulates prediction as estimating a patient’s risk from the patient’s historical, time-ordered EMR records.
- ESRD patients face severe life threats, lifelong treatment, and periodic hospital visits for multifarious tests, motivating dynamic health-risk prediction.
- Given historical EMR records from a patient’s visits, the model predicts the probability of a specific risk such as mortality, disease diagnosis, or decompensation.
Methods
AdaCare combines multi-scale dilated convolutions, raw and historical features, and adaptive feature recalibration within a GRU-based architecture to represent visit-level health status and predict risk.
- Visit-level representation: AdaCare uses a GRU-based architecture that combines the latest record with historical characteristics to build each visit embedding and predict healthcare status.Using both inputs balances sensitivity to recent abnormalities with information from longer-term patient history.
- Multi-Scale Dilated Convolution: Multi-scale dilated convolution captures long-term biomarker trends and short-term abnormal variations through parallel branches with different dilation rates.Causal padding prevents follow-up records from leaking into the extracted historical features.
- Multi-Scale Dilated Convolution: Dilation rate k = 1 examines consecutive records, while k = 2 and k = 3 skip one and two records per input, respectively.These rates provide short-, medium-, and relatively long-term temporal receptive fields.
- Scale-Adaptive Clinical Feature Recalibration: A scale-adaptive feature recalibration module learns feature weights and applies them through element-wise multiplication to reduce redundancy in the clinical representation.The recalibrated features are adjusted dynamically across visits according to the patient’s health condition.
- Visit-level representation: AdaCare concatenates weighted raw features with weighted convolutional features before feeding the visit embedding into a GRU for sequential health-status representation.The resulting selectively enhanced predictive features can serve as a precursor of health risk for the given patient.
Experiment
AdaCare is evaluated on decompensation prediction using MIMIC-III and mortality prediction using ESRD data, with comparisons against baseline models and variants. Results assess predictive performance, dynamic feature recalibration, and interpretability through feature-importance analysis and clinical knowledge review.
- Experimental Setup: AdaCare is evaluated on decompensation prediction with MIMIC-III data and mortality prediction with ESRD data.The experiments use two real-world clinical datasets and report AUPRC, Min(Se,P+), and AUROC.
- Experimental Setup: The comparison includes GRU, RETAIN, T-LSTM, SAnD, and AdaCare variants with multi-scale convolution and feature recalibration.The variants distinguish multi-scale dilated convolution and sigmoid or sparsemax recalibration modules.
- Prediction Results: AdaCare outperforms all baseline models across both datasets and all evaluation metrics, with AUPRC improvements of 14.2% on ESRD and 9.1% on MIMIC-III over the best baseline.AUPRC is emphasized as the primary metric for highly imbalanced EMR data.
- Ablation Results: The multi-scale convolution variant outperforms the baselines, while adding sigmoid-based feature recalibration further improves performance.These comparisons support the use of temporal variation patterns and feature recalibration in the representation.
- Ablation Results: Sparsemax recalibration enhances only a few predictive features and performs slightly worse than sigmoid recalibration on ESRD mortality prediction.The sparsemax variant is examined for emphasizing the most high-risk clinical features at the latest visit.
- Interpretability Analysis: Average biomarker-importance analysis links serum albumin, urea, and serum chlorine to specific adverse outcomes and disease-related conditions in ESRD patients.The analysis uses ESRD validation sets and summarizes medical knowledge learned from feature recalibration.
Conclusion
AdaCare learns clinical health-status representations by combining multi-scale temporal feature extraction with nonlinear feature recalibration. Experiments on MIMIC-III and ESRD show stronger performance than baseline approaches, while its learned medical knowledge is supported by experts and related literature.
- Conclusion: AdaCare uses multi-scale dilated convolutions to capture long- and short-term biomarker variation patterns.It extends a squeeze-and-excitation block to model nonlinear feature dependencies and selectively enhance predictive features.
- Conclusion: AdaCare outperforms baseline approaches on MIMIC-III and ESRD while providing interpretability.Medical knowledge learned by AdaCare was positively confirmed by human medical experts and related medical literature.