Abnormality prediction and forecasting of laboratory values from electrocardiogram signals using multimodal deep...

Area of Science:

• Biomedical Engineering
• Medical Informatics
• Artificial Intelligence in Medicine

Background:

• Laboratory tests are crucial for diagnosing and monitoring various health conditions.
• Current methods for lab testing can be invasive, costly, and time-consuming.
• There is a need for non-invasive and efficient methods to predict and monitor laboratory abnormalities.

Purpose of the Study:

• To investigate the feasibility of using electrocardiogram (ECG) data and patient metadata to predict and forecast laboratory abnormalities.
• To develop and evaluate multimodal deep learning models for this task.
• To assess the potential of this approach as a non-invasive, cost-effective alternative to traditional lab testing.

Main Methods:

• Utilized the MIMIC-IV dataset for training multimodal deep learning models.
• Incorporated ECG waveforms, demographics, biometrics, and vital signs as input features.
• Employed a structured state space classifier with late fusion for metadata integration.
• Framed the task as individual binary classifications for each abnormality and evaluated using AUROC.

Main Results:

• Achieved strong performance in predicting and forecasting 24 different lab values across various categories (cardiac, renal, hematological, etc.).
• AUROCs exceeded 0.70 for abnormality prediction and forecasting.
• NTproBNP prediction achieved an AUROC > 0.90.
• Hemoglobin, Albumin, and Hematocrit also showed high prediction accuracy (AUROC > 0.85).

Conclusions:

• ECG data combined with clinical data can effectively predict and forecast laboratory abnormalities.
• This multimodal approach offers a non-invasive and cost-effective solution for patient monitoring.
• The findings support early intervention and improved patient care by reducing reliance on traditional lab testing.