Medication Recommender System for ICU Patients Using Autoencoders.

Area of Science:

• Medical Informatics
• Clinical Pharmacology
• Artificial Intelligence in Medicine

Background:

• Intensive care units (ICUs) frequently involve high-risk medications, increasing the potential for prescribing errors.
• Overprescribing, underprescribing, and inappropriate medication choices are significant risks in the ICU setting.

Purpose of the Study:

• To evaluate recommendation methods for suggesting appropriate ICU medications.
• To determine if incorporating clinical patient data improves medication recommendations.

Main Methods:

• Formulated medication list completion as a recommendation task using the MIMIC-III dataset.
• Analyzed four autoencoder-based approaches and two baseline methods.
• Input data included either only known medications or medications combined with patient data.

Main Results:

• Autoencoder-based medication recommender systems demonstrated success in recommending medications for ICU patients.
• Incorporating clinical patient data alongside medication lists potentially enhances recommendation accuracy.

Conclusions:

• Medication recommender systems show promise for improving medication safety in ICUs.
• Autoencoder models are effective for medication recommendation tasks, especially when enriched with patient clinical data.