From data to optimal decision making: a data-driven, probabilistic machine learning approach to decision support for...

Area of Science:

• Medical Informatics
• Clinical Decision Support
• Sepsis Management

Background:

• Challenges exist in extracting actionable knowledge from heterogeneous electronic health records (EHR) for clinical decision support.
• Limited analytical tools hinder efficient extraction of informed decisions due to data complexity and missing information.

Purpose of the Study:

• To develop and assess a data-driven method for inferring patient states, predicting sepsis trajectories, and optimizing antibiotic administration.
• To predict mortality and length of stay for sepsis patients using a probabilistic framework.

Main Methods:

• A Partially Observable Markov Decision Process (POMDP) model was developed using EHR data from 1492 sepsis patients.
• The framework defines states, actions, and rewards based on clinical practice and expert knowledge.
• Model performance was evaluated on a separate test set, focusing on antibiotic administration policies and mortality/length-of-stay prediction.

Main Results:

• Data-derived antibiotic policies improved favorable patient outcomes by 49% compared to 37% with alternative policies (P=1.3e-13).
• The model demonstrated robustness against parameter variations and data uncertainty.
• Mortality prediction achieved an AUC of 0.7, and length-of-stay prediction showed similar performance (AUC 0.69-0.73).

Conclusions:

• A data-driven model effectively suggests favorable treatment actions and accurately predicts sepsis patient outcomes.
• This framework provides a scalable solution for probabilistic clinical decision support in sepsis and can be adapted to other clinical areas.