Abstract
Objective: Leveraging patient data through machine learning techniques in disease care offers a multitude of substantial benefits. Nonetheless, the inherent nature of patient data poses several challenges. Prevalent cases amass substantial longitudinal data owing to their patient volume and consistent follow-ups, however, longitudinal laboratory data are renowned for their irregularity, temporality, absenteeism, and sparsity; In contrast, recruitment for rare or specific cases is often constrained due to their limited patient size and episodic observations. This study employed self-supervised learning (SSL) to pretrain a generalized laboratory progress (GLP) model that captures the overall progression of common laboratory markers in prevalent cardiovascular cases, with the intention of transferring this knowledge to aid in the detection of specific cardiovascular event. Methods and procedures: GLP implemented a two-stage training approach, leveraging the information embedded within interpolated data and amplify the performance of SSL. After GLP pretraining, it is transferred for TVR detection. Results: The proposed two-stage training improved the performance of pure SSL, and the transferability of GLP exhibited distinctiveness. After GLP processing, the classification exhibited a notable enhancement, with averaged accuracy rising from 0.63 to 0.90. All evaluated metrics demonstrated substantial superiority ($p < 0.01$) compared to prior GLP processing. Conclusion: Our study effectively engages in translational engineering by transferring patient progression of cardiovascular laboratory parameters from one patient group to another, transcending the limitations of data availability. The transferability of disease progression optimized the strategies of examinations and treatments, and improves patient prognosis while using commonly available laboratory parameters.
Introduction
Electronic Health Record (EHR) data is notoriously difficult to work with. Longitudinal laboratory results are often irregular (tests happen at random intervals), sparse (long gaps between visits), and incomplete. While massive amounts of data exist for prevalent conditions like hypertension (HTN) or diabetes (DM), specific high-risk events like Target Vessel Revascularization (TVR)—a complication where a previously treated heart vessel needs another procedure—often suffer from limited data availability. Training deep learning models on these small, "episodic" datasets usually leads to overfitting. This study introduces a Generalized Laboratory Progress (GLP) model that uses self-supervised learning to bridge this gap by transferring knowledge of disease progression from large populations to specific cardiovascular event detection.
Methodology
The GLP model utilizes a two-stage training approach to handle the challenges of irregular and sparse clinical data. The core architecture features a Longitudinal Iterative Block (LIBC) consisting of a Bidirectional LSTM (BiLSTM) to capture temporal context, followed by a regressor.
Two-Stage Training
- Stage 1 (Interpolation Training): To address irregular timing, the model uses three interpolation methods (Linear, PCHIP, and Barycentric) to create "interpolated frames." The model is trained in a supervised manner to predict the next month's lab values based on these smoothed trends.
- Stage 2 (Self-Supervised Refinement): The model then employs an autoregressive self-supervised learning (SSL) approach on the original, non-interpolated data. It learns to predict future lab values based on its own previous generations, refining its internal representation of "disease progression."
Results and Discussion
The GLP model was pretrained on a dataset of 9,720 patients with HTN/DM and evaluated on a downstream dataset of 483 patients who underwent percutaneous coronary intervention (PCI). The impact of using the GLP-processed data versus raw clinical data was dramatic:
- Accuracy Boost: Average classification accuracy for TVR detection rose from 0.63 to 0.90.
- AUROC: Reached 0.91 using the GLP-processed features, compared to much lower scores using original data (~0.51).
- Transferability of Progression: The most significant finding is that the pattern of how lab markers change is transferable across different patient cohorts. Even though the pretraining was on HTN/DM patients, the model learned a "biological language" of lab trends highly relevant to post-surgical cardiac patients.
- The "Distillation" Effect: The GLP model essentially acts as a noise filter and feature extractor that "distills" messy lab history into a clear trajectory, making it much easier for downstream classifiers to identify risk.
Conclusion and Clinical Impact
This study effectively transposes patient progression from one cohort to another, surpassing the constraints of episodic observation. The transferability of disease progression optimized the strategies of examinations and treatments, improving patient prognosis while using commonly available laboratory parameters. Future work may expand this framework to other rare diseases where longitudinal data is scarce but common lab markers are available.
Summary updated