Modelling patient trajectories using multimodal information.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Computational Biology

Background:

• Electronic Health Records (EHRs) contain longitudinal patient data crucial for health monitoring and prognosis.
• The complexity of EHR data, including long time spans and rapid data generation, challenges clinical decision-making.
• Patient trajectory modeling offers a scalable approach to leverage EHR data for improved healthcare quality and preventive medicine.

Purpose of the Study:

• To propose and evaluate flexible deep learning architectures for modeling patient trajectories.
• To integrate diverse data types, including clinical text and standard codes, while considering the temporal nature of data.
• To enhance the accuracy of clinical predictions such as disease progression and patient readmission.

Main Methods:

• Developed two distinct deep learning architectures for patient trajectory modeling.
• The first architecture converts patient admissions into dense representations for flexible feature sets.
• The second architecture employs a recurrent-based model with a sliding window mechanism to process admission representations sequentially.

Main Results:

• Evaluated models on the MIMIC-III database for predicting unexpected patient readmission and disease progression.
• The first architecture showed potential for readmission and diagnosis prediction based on single admissions.
• The sequence-based architecture achieved performance comparable to existing solutions, demonstrating the utility of the sliding window approach.

Conclusions:

• Explored deep learning techniques for patient trajectory modeling using individual and sequential analysis of admissions.
• Combining clinical text with other data types yielded positive results, with potential for further improvement using fine-tuned clinicalBERT.
• The developed patient trajectory modeling solution is publicly available for research and application.