AI-based modeling of treatment decisions in benign prostatic hyperplasia: a transformer-based comparative study

Comprehensive Summary

Alshraideh et al. examined whether transformer-based machine learning models have the ability to predict treatment decisions in benign prostatic hyperplasia (BPH). Their task was to determine whether the patients should undergo surgery (TURP) or be on medical therapy. The analysis consisted of 883 patients data collected between 2017 and 2024. The authors compared several models, including GEMMA and GPT (transformer-based models), as well as LSTM, CNN, RNN, which were for baseline comparison. The results showed that transformer models performed best, with GEMMA reaching 92% accuracy and an AUC of 0.94, while GPT achieved 91% accuracy. Other models performed within the range of 90% to 84%, which conveys that more complex models capture clinical relationships better. Further analysis showed thatPSA levels, prostate size, urinary retention, and creatinine were the strongest predictors. These models produced stable predictions and could offer some clinical value beyond simple heuristics. However, it is notable that the models were trained on past physician decisions, meaning it reflects practice patterns rather than true treatment effectiveness.

Outcomes and Implications

This study shows that machine learning models are capable of organizing complex clinical data into structured predictions that may support treatment decisions in BPH. Compared to standard clinical judgment alone, these models can compare various variables at once and produce more consistent outputs. This could be useful in settings where treatment decisions vary between clinicians. However, given that the data is based on a single institution, performance in other settings is uncertain. There is also a risk that the model reinforces existing biases in physician decision-making instead of improving them. For now, this type of model works best as a decision-support tool rather than something that replaces clinical judgment.