Seminar "Evaluation of fall risk factors from a real-world dataset using Bayesian networks"

Summary:

Falls are a significant problem for older adults, and identifying and assessing risk factors is essential to reducing fall rates. However, fall prevention requires an educational and repeated approach, time, and expertise to accurately target actionable risk factors. This thesis aims to assess fall risk factors using a real-world dataset and Bayesian networks.

The use of real data poses challenges, particularly with respect to data preprocessing, which is time consuming and requires expertise. In addition, an AI-based application raises new challenges such as trust, which depends on the interpretability and explainability of results. To address these challenges, this thesis proposes a knowledge model (Bayesian networks) that automatically evaluates key actionable risk factors. The model is trained on a real-world data set combined with expert knowledge. Two iterations of the data preprocessing steps are presented and explained, including imputation of missing values, variable selection, and use of balancing techniques for unbalanced data. The first iteration included only the main variables to validate the feasibility of the process, and the second iteration included as many variables as possible to improve prediction and the process.

The model is compared to other well-known classifiers through a variety of measures, including full or partial observation, and whether or not balancing methods are used to handle the tricky issue of unbalanced data. A Bayesian network is presented as a good solution, combining the quality of the results to assess risk factors and the interpretability/explicability of the model from the expert's perspective.
The results show that predicting the presence or absence of fall risk factors is a difficult task. While Bayesian networks and other classifiers have equivalent performance in terms of measures such as balanced accuracy and f1 score, the value of Bayesian networks lies in their interpretability and the ability to use partial observations.

In summary, this thesis presents a contribution to an application for falls prevention that facilitates automatic assessment of risk factors from partial patient observations, using a real-world data set and Bayesian networks. The proposed knowledge model (Bayesian networks) addresses the challenges of using real data and AI-based applications, respectively.