A Longitudinal Patient Preference-Aware Treatment Recommendation System Using Multi-Source Sentiment Fusion and Adaptive Utility Modelling

Abstract

Treatment recommendation systems using artificial intelligence have shown significant gains in predictive accuracy and clinical decision support, but an inherent design challenge exists in virtually all existing frameworks, namely that patient preferences are either considered as fixed binary inputs, they are viewed as auxiliary features, or are not part of the optimization problem at all. There are severe clinical implications of such a gap, with guidelines that do not incorporate changing patient values being always linked with declining therapeutic adherence, diminishing patient contentment, and structural devaluation of shared decision-making, which is progressively implemented as a fundamental principle of ethical clinical care by regulatory authorities such as the FDA, the European AI Act, and the American Society of Clinical Oncology. To overcome this daunting constraint, this paper presents PreferenceNet, a longitudinal, preference-sensitive, treatment recommendation system, which can model, track, and optimize against patient preferences during repeated clinical visits over time. The system operates through four tightly integrated modules: a Multi-Source Preference Signal Extraction pipeline fusing patient-reported outcome surveys, clinical consultation transcripts processed via fine-tuned BioBERT, and wearable behavioral time-series encoded through a bidirectional LSTM into a unified preference representation; a Temporal Preference Drift Model employing a Milestone-Conditioned Hidden Markov Model that infers latent preference states and detects clinically meaningful transitions across longitudinal visit sequences from MIMIC-IV; an Adaptive Utility Optimization Layer with patient-specific, time-varying weights across clinical efficacy, preference alignment, side-effect tolerance, and treatment burden, updated via Bayesian inference at each clinical encounter; and a Preference-Concordant Recommendation module quantifying outputs through two originally defined metrics — the Preference Concordance Score and Utility Alignment Index — delivered via a clinician-facing interpretability dashboard. Assessed on 12,847 patients in comparison to four fixed baselines, PreferenceNet shows an AUC of 0.87, Preference Concordance Score of 0.79, Utility Alignment Index of 0.81, and Preference Drift Detection Accuracy of 83.2 percent, and an 88 percent clinical agreement rate under an additional vignette study, which together indicate.