AI-Driven Prediction of Anti-inflammatory Dose Requirements and Treatment Response for Patient-Specific Management of Ulcerative Colitis

Background

Ulcerative colitis (UC) is a chronic inflammatory bowel disease characterized by patient heterogeneity and variable therapeutic response. Current treatments remain empirical, while drug repurposing offers a promising strategy to accelerate therapeutic development. Vardenafil, a phosphodiesterase-5 inhibitor with anti-inflammatory potential in UC, represents a model repurposed drug.

Aim
This study aimed to develop AI-based predictive models for patient-specific management of UC, using vardenafil as a model repurposed drug.

Methods
A dataset of 370 patient records retrospectively curated from published UC studies involving different anti-inflammatory therapies. Animal data were converted to human-equivalent data, where applicable. Patient-specific demographic and clinical characteristics, including age, weight, disease duration, smoking status, family history, Mayo score, C-reactive protein, interleukin-6, and fecal calprotectin, were used as predictive variables. Three machine learning models predicted dose requirement (mg/day), treatment response (0-100%), and the subsequent intervention (no action, dose escalation, or surgery). Several machine learning algorithms were tested to select the ones with the least errors. Models were developed using a 70:30 training/testing split, with cross-validation during hyperparameter optimization. Vardenafil was selected as the model repurposed drug, based on preclinical evidence and model-derived predictions were evaluated in a rat model of UC.

Results
Gradient Boosting demonstrated the strongest performance among algorithms, for the three models. Dose prediction achieved R² = 0.984 and RMSE = 0.121 while treatment-response prediction achieved R² = 0.730 and RMSE = 12.976. Prediction of intervention achieved an F1-score of 0.845, precision of 0.834, and recall of 0.868. Experimental evaluation showed close agreement between model-derived and observed outcomes, providing preliminary support for predicted treatment effectiveness.

Conclusion
AI-driven modelling provides a promising approach for patient-specific prediction of anti-inflammatory dose requirements, treatment response and subsequent intervention in UC.  However, prospective clinical validation is required before clinical implementation and to determine whether the approach improves patient outcomes.