Lay Abstract for Statistics in Medicine Article: Relative sparsity for medical decision problems

Each week, we publish lay abstracts of new articles from our prestigious portfolio of journals in statistics. The aim is to highlight the latest research to a broader audience in an accessible format.

The article featured today is from Statistics in Medicine with the full article now available to read here.

Weisenthal, SJThurston, SWErtefaie, ARelative sparsity for medical decision problemsStatistics in Medicine202342(18): 30673092. doi: 10.1002/sim.9755

There is much interest in the statistics and machine learning communities in developing models to help healthcare providers and patients make medical decisions.  Healthcare providers are often tasked with making challenging, life-altering decisions, and decision aids can be useful. Models for making medical decisions can be trained on large amounts of data and may consider many different patient characteristics, such as vital signs and laboratory values.  

In the proposed work, the authors extend an algorithm called Trust Region Policy Optimization (TRPO), which was designed for robotics applications, to the medical setting.  In the medical setting, it is important to explain and justify a new change in a treatment strategy to interested parties, such as the healthcare provider and patient. Justification of a new treatment model is facilitated if the model has a succinct, interpretable difference from the established care guidelines, or the standard of care.    

The authors therefore propose a method to select models that differ from the standard of care only with respect to how they consider a small number of patient characteristics. They call this new method “relative sparsity.” For example, both the standard of care and the new treatment model may consider many patient characteristics, but the strategy for the new treatment model might differ from the standard only in its emphasis on heart rate.  

The authors explore the properties of their method with simulated data. They also analyze a real, observational, electronic health record dataset, and a model is obtained to help make decisions in the setting of dangerously low blood pressure in the intensive care unit. The new model, which improves final blood pressure readings, considers many patient characteristics but only differs from the standard of care with respect to the emphasis it puts on the patient’s initial blood pressure.  

In its transparency, the proposed method facilitates the translation of data-driven decision aids from the research bench to the patient bedside, where these decision aids have exciting potential to improve outcomes. 

More Details