Healthcare applications of 0-1 neural networks in prescriptive problems with observational data

Area of Science:

• Artificial Intelligence in Medicine
• Machine Learning for Healthcare
• Personalized Medicine

Background:

• Medical decision-making with limited observational data is challenging, especially in personalized healthcare.
• Existing models struggle to balance interpretability and policy complexity.
• The intricate relationships between patient characteristics, treatments, and outcomes require advanced modeling.

Purpose of the Study:

• Introduce prescriptive neural networks (PNNs) for optimizing treatment policies in medium-data settings.
• Enhance interpretability and policy complexity compared to deep neural networks and decision trees.
• Evaluate PNNs' performance against existing methods in synthetic and real-world healthcare data.

Main Methods:

• Developed shallow 0-1 neural networks (PNNs) trained with mixed integer programming.
• Utilized counterfactual estimation for policy optimization.
• Compared PNNs with existing methods on synthetic datasets and a postpartum hypertension treatment case study.

Main Results:

• PNNs demonstrated superior performance in optimizing treatment policies.
• In a postpartum hypertension case study, PNNs reduced peak blood pressure by 5.47 mm Hg (p=0.02) compared to clinical practice.
• PNNs outperformed the next best prescriptive modeling technique by 2 mm Hg (p=0.01) and were more likely to identify clinically significant features, avoiding biased ones.

Conclusions:

• Prescriptive neural networks (PNNs) offer a powerful, interpretable approach for medical decision-making with limited data.
• PNNs can lead to improved patient outcomes and more equitable healthcare by avoiding reliance on biased features.
• The study highlights the potential of PNNs to advance personalized medicine and optimize clinical practice.