Introduction
This project expands upon the paper Benchmarking Deep Learning Architectures for Predicting Readmission to the ICU and Describing Patients-at-Risk, which utilizes a variety of neural network architectures to predict the probability of patient readmission to an ICU using prior patient medical data. Predicting patient readmission allows doctors to craft tailored treatment plans and allocate hospital resources accordingly.
Models using attention mechanisms, recurrent layers, neural ordinary differential equations (ODEs), and medical concept embeddings with time-aware attention were trained on the MIMIC-III dataset, comprising 45,298 ICU stays for 33,150 patients.
Data Description
The MIMIC-III dataset contains longitudinal electronic medical records for numerous patients. Two distinct inputs are fed into the model: static and timestamped codes.
- Static Data: Variables such as age, sex, insurance type, and marital status.
- Timestamped Codes: Medical data like diagnoses, prescriptions, and procedures, providing a patient’s medical history.
Only ICU admission data is used, ensuring compliance with patient anonymity and medical privacy laws.
Dataset Features
- Clinical Notes
- Demographic Information
- Admission and Discharge Information
- ICD-9 Codes
- Lab Results
- Medications
- Vital Signs
- Procedures Performed
- Imaging Data
Results
Comparison of original and our results:
Resources
Conclusion
This project benchmarks multiple deep learning architectures for ICU readmission prediction, enabling better resource allocation and patient care. By leveraging the MIMIC-III dataset, it advances the understanding of risk factors for ICU readmissions.