Focal therapy is an increasingly utilised prostate cancer treatment option, although there are limitations to current diagnostic tools for FT patient selection. Integral to clinical decision making is accurate cancer detection, diagnosis and staging. Given large volumes of clinical data, it is now feasible to build Artificial Intelligence (AI) classification models to support and augment patient selection for different treatment types. This AI adjunct can improve decision-making accuracy and address potential clinician biases. The aim of this study is to build and evaluate an AI classification model to aid patient selection for prostate cancer focal therapy.

Clinical data including patient baseline characteristics, imaging findings from MRI and PSMA PET/CT, and histopathology results from 231 men who underwent radical prostatectomy for prostate cancer treatment between April 2015 and April 2023 were collected. The data were pre-processed using standardized normalization techniques and then integrated to build several AI classification models for predicting eligibility for FT. The machine learning algorithms used for building these AI models were random forest, gradient boost, support vector machines, and logistic regression, all of which were designed towards a binary classification outcome. After the models were trained and validated, they were evaluated using ROC analysis on test sample sets comparing its performance against current selection criteria and gold-standard outcomes to assess patient selection accuracy.

All models demonstrated superior performance compared to conventional selection criteria based on standard clinical and imaging assessments. The highest performance was the random forest model, gradient boosting and XG boost with an AUC of 1. When compared against gold-standard outcomes for FT eligibility, out of the best models, after hyperparameter tuning, the random forest model achieved an AUC of 0.91, with a sensitivity of 0.5 and a specificity of 0.88. Variable analysis indicated that the absence of contralateral disease was the most predictive feature for FT eligibility, with Gleason grade groups from transperineal prostate biopsy as the next most significant factor.

Accurate FT patient selection is critical for treatment success. In patients already selected for other treatment like whole-gland options, this AI classification model can assist as an unbiased adjunct in confirming exclusion for FT. The effectiveness of this model further suggests that AI has a potential role in integrating clinical data to result in a singular prediction for prostate cancer treatment. Further external validation and prospective studies are required to confirm the widespread clinical utility of AI classification models.