AI Helper for Clearer Fall-Risk Models

Falls and related injuries matter for people 50 and older. A study asks whether AI can help build clearer risk models. The aim is not to replace simple models. The aim is to make them work better with lots of data.

Cox regression is a common transparent tool for health risk estimates. People can read it and trust it. But with dozens or even hundreds of inputs, hand-building the best version gets hard. Someone must choose inputs, reshape them, and check how they combine. Complex machine learning can score well. Yet its opaque style can make clinicians uneasy. Trust, transparency, and interpretability matter for decisions.

The project built an explainable AI recommender. It gives data-driven suggestions about features. Those suggestions can be folded into an existing readable statistical model. The test case was a Cox model for the one-year risk of a first fall or related injury. The data came from linked primary and secondary care records in DataLoch, Lothian, Scotland. It covered adults 50 and older who were registered with a contributing general practice from January 2015 to February 2020.

A random survival forest was explained with SHAP. Its feature attributions became recommendations for the Cox model. The suggestions covered leaving features out, adding nonlinear terms, and adding interactions. The data were split 80/20. The work included internal validation. The point is to ask whether AI guidance can improve a model people can still understand. That question matters. A model that performs well but cannot be explained may not earn trust. A model that is clear but poorly built may miss useful patterns.