A Learnable Multi-views Contrastive Framework with Reconstruction Discrepancy for Medical Time-Series

TL;DR

The paper tackles annotation scarcity and cross-center heterogeneity in medical time-series diagnosis by coupling cross-center knowledge transfer via AE-GAN with a Learnable Multi-Views Contrastive Framework. It introduces a two-stage pipeline where external healthy data informs feature augmentation through reconstruction discrepancies, followed by adaptive, multi-view contrastive learning that leverages subject, trial, epoch, and temporal structure with inter-view and intra-view losses. Empirically, LMCRD achieves state-of-the-art performance across myocardial infarction, Alzheimer’s disease, and Parkinson’s disease datasets, maintaining strong results even with limited labeled data. The approach yields informative, multi-view representations that improve generalization and offer practical benefits for healthcare diagnostics and time-series analysis.

Abstract

In medical time series disease diagnosis, two key challenges are identified.First, the high annotation cost of medical data leads to overfitting in models trained on label-limited, single-center datasets. To address this, we propose incorporating external data from related tasks and leveraging AE-GAN to extract prior knowledge,providing valuable references for downstream tasks. Second, many existing studies employ contrastive learning to derive more generalized medical sequence representations for diagnostic tasks, usually relying on manually designed diverse positive and negative sample pairs.However, these approaches are complex, lack generalizability, and fail to adaptively capture disease-specific features across different conditions.To overcome this, we introduce LMCF (Learnable Multi-views Contrastive Framework), a framework that integrates a multi-head attention mechanism and adaptively learns representations from different views through inter-view and intra-view contrastive learning strategies.Additionally, the pre-trained AE-GAN is used to reconstruct discrepancies in the target data as disease probabilities, which are then integrated into the contrastive learning process.Experiments on three target datasets demonstrate that our method consistently outperforms seven other baselines, highlighting its significant impact on healthcare applications such as the diagnosis of myocardial infarction, Alzheimer's disease, and Parkinson's disease.

Figures (6)

• Figure 1: Training Procedure of LMCRD
• Figure 2: Mechanism of Multi-View Contrastive Learning
• Figure 3: Visualization of Multi-View Representations on the AD Dataset. The figure compares contrastive view representations obtained through self-supervised contrastive learning and 100% supervised fine-tuning.
• Figure B1: Structure of Medical Time Series.The data is organized into four levels: Subject ($p_j$), Trial ($r_k$), Epoch ($e_m$), and Temporal ($e_{m,t}$), capturing different granularities of the data, from subject-level recordings to temporal observations within individual epochs.
• Figure B2: Overview of LMCRD