RL-Guided Adaptive Dose Optimization for Chemotherapy

The Landscape

Cancer treatment is one of the most consequential sequential decision-making problems in medicine. A patient undergoing chemotherapy receives a series of treatment cycles -- typically 4 to 8 rounds spaced 2 to 3 weeks apart. At each cycle, the oncologist must decide: what dose should the patient receive?

Too high a dose causes severe side effects -- neutropenia, organ damage, treatment discontinuation. Too low a dose allows the tumor to grow back between cycles. The optimal dosing strategy depends on the patient's current health state: tumor size, white blood cell count, kidney function, and overall toxicity level.

Traditional clinical practice uses fixed dosing protocols. A patient weighing 70 kg with a certain body surface area receives a standard dose, regardless of how their body is responding. Some oncologists adjust doses reactively -- reducing the dose after severe side effects appear -- but this is ad hoc and inconsistent.

NovaCure Therapeutics

NovaCure Therapeutics is a mid-stage pharmaceutical company with $420M in annual revenue, focused on solid tumor oncology. They have a portfolio of three chemotherapy drugs in clinical trials and two FDA-approved treatments.

Their clinical data spans 12,000 patients across 8 clinical trials, with detailed longitudinal records: tumor measurements via imaging (every 3 weeks), complete blood counts (weekly), liver and kidney panels (biweekly), and patient-reported quality-of-life scores.

The Problem

NovaCure's lead drug, NC-4817, shows promising efficacy in Phase II trials for advanced non-small-cell lung cancer. However, the trial data reveals a critical pattern:

• 34% of patients require dose reductions due to toxicity
• 18% discontinue treatment entirely due to adverse effects
• Among patients who complete all cycles, tumor response varies enormously -- some achieve complete remission, others show minimal response

The oncology team suspects that a personalized, adaptive dosing strategy could simultaneously improve efficacy (better tumor control) and reduce toxicity (fewer dose reductions and discontinuations). But the challenge is that the optimal dose at cycle $k$ depends on everything that has happened in cycles 1 through $k-1$.

This is a sequential decision-making problem under uncertainty. It is exactly the kind of problem that reinforcement learning was designed to solve.

Business Impact

If NovaCure can demonstrate that an RL-guided dosing protocol improves outcomes, the potential impact is:

• Regulatory: A companion dosing algorithm submitted alongside the drug could accelerate FDA approval and differentiate NC-4817 from competitors
• Clinical: Improved patient outcomes (higher response rates, fewer adverse events) directly impact the drug's commercial value
• Financial: Reducing treatment discontinuation from 18% to under 10% could increase revenue by $45M annually for NC-4817 alone
• Platform: A validated RL dosing framework could be applied across NovaCure's entire drug portfolio

---