Adaptive Dose Optimization Algorithm for LED-based Photodynamic Therapy Based on Deep Reinforcement Learning

Abstract

Current photodynamic therapy fails. Not occasionally-systematically. We measured optical coefficients across 1,847 lesions: peripheral zones exhibit μs' ≈ 20 cm⁻¹ while necrotic zones μs' ≈ 0.8 cm⁻¹, a 25-fold discontinuity that renders uniform protocols obsolete. Drug clearance varies threefold. One patient metabolizes protoporphyrin IX in two hours; another requires six. Oxygen maps tell an equally chaotic story-partial pressures crash from 95 mmHg to anoxic thresholds within minutes, creating dead zones where photochemistry simply stops. We built a system that adapts. Deep reinforcement learning processes 384 physiological signals in real-time, adjusting power density and fractionation schedules every 100 milliseconds. The architecture splits decision-making: one network learns patient baselines, another computes action advantages. This dueling structure stabilizes training on sparse clinical data, where 62.5% of treatments historically achieved only partial or no response. Results contradict decades of conservative practice. Phototoxicity decreased from 18.3% to 7.6%, a relative risk reduction of 58% (95% CI: 49-66%). Complete responses jumped from 37.8% to 58.7%-not through gentler treatment, but through aggressive, precisely-timed interventions the algorithm discovered autonomously. A paradox emerged: lower doses for superficial lesions, intense protocols for deep tumors, opposite to clinical intuition. The system runs on standard GPUs. No specialized hardware. Forty-seven prospective patients confirmed what retrospective analysis suggested: adaptive control fundamentally outperforms fixed protocols.