Abnormality Prediction and Forecasting of Laboratory Values from Electrocardiogram Signals Using Multimodal Deep Learning

TL;DR

This study investigates predicting laboratory abnormalities from ECG signals combined with basic patient metadata using a multimodal deep learning framework. It frames the problem as dual tasks: abnormality prediction and abnormality forecasting, leveraging a S4-based time-series encoder for ECG features and an MLP for static features, with late fusion of modalities on the MIMIC-IV data. The results show robust performance, with AUROCs exceeding 0.70 for 24 abnormalities and NT-proBNP predictions reaching AUROCs around 0.90, demonstrating potential for non-invasive, cost-effective monitoring and early intervention. While promising, the work notes limitations related to threshold derivation, generalizability, and the need for external validation and explainability before clinical deployment.

Abstract

This study investigates the feasibility of using electrocardiogram (ECG) data combined with basic patient metadata to estimate and monitor prompt laboratory abnormalities. We use the MIMIC-IV dataset to train multimodal deep learning models on ECG waveforms, demographics, biometrics, and vital signs. Our model is a structured state space classifier with late fusion for metadata. We frame the task as individual binary classifications per abnormality and evaluate performance using AUROC. The models achieve strong performance, with AUROCs above 0.70 for 24 lab values in abnormality prediction and up to 24 in abnormality forecasting, across cardiac, renal, hematological, metabolic, immunological, and coagulation categories. NTproBNP (>353 pg/mL) is best predicted (AUROC > 0.90). Other values with AUROC > 0.85 include Hemoglobin (>17.5 g/dL), Albumin (>5.2 g/dL), and Hematocrit (>51%). Our findings show ECG combined with clinical data enables prompt abnormality prediction and forecasting of lab abnormalities, offering a non-invasive, cost-effective alternative to traditional testing. This can support early intervention and enhanced patient monitoring. ECG and clinical data can help estimate and monitor abnormal lab values, potentially improving care while reducing reliance on invasive and costly procedures.

Figures (2)

• Figure 1: Schematic overview of the study incluiding predictive workflow and task definitions. A) Illustration of the predictive workflow used in the study where model inputs include ECG waveforms, demographics, biometrics, and vital signs, which are used in a binary classification setting to predict abnormal laboratory values. B) Demonstration of the abnormality prediction task, where we sample the closest vital signs within 30 minutes around the ECG record and use the closest laboratory value within 60 minutes as prediction target. C) Demonstration of the abnormality forecasting task, where we also use closest vital signs within 30 minutes around the ECG record as input but where we predict the presence of any abnormal laboratory value within a predefined time horizon.
• Figure 2: Schematic overview of the dataset creation process, starting from three source datasets (MIMIC-IV-ECG, MIMIC-IV, and MIMIC-IV-ECG-ICD) and applying various inclusion criteria. The process includes patient selection equal or above 18 years old, unit standardization for laboratory values, outlier removal, and label definition based on median-derived thresholds. Similarly, we incorporate a special token in the label space for labels that were not available to sample for that specific sample, these tokens help as placeholders to be excluded during training and evaluation which ensure large availability of samples and labels even when some labels are not available for a sample. Additionally, we incorporated an imputation strategy to handle missing data with the use of binary masks as helpers e.g. one binary feature per each feature to let the model know which values are real and which imputed. Furthermore, we selected only labels with at least 20 available (10 positive and 10 negative) counts. Finally, we dropped samples that does not contain any available labels after the already mentioned steps.