de Saporta Benoîte, Thierry d'Argenlieu Aymar, Sabbadin Régis, Cleynen Alice
IMAG, CNRS, Univ Montpellier, Montpellier, France.
IP Paris, Palaiseau, France.
PLoS One. 2024 Dec 19;19(12):e0315661. doi: 10.1371/journal.pone.0315661. eCollection 2024.
Designing patient-specific follow-up strategies is key to personalized cancer care. Tools to assist doctors in treatment decisions and scheduling follow-ups based on patient preferences and medical data would be highly beneficial. These tools should incorporate realistic models of disease progression under treatment, multi-objective optimization of treatment strategies, and efficient algorithms to personalize follow-ups by considering patient history. We propose modeling cancer evolution using a Piecewise Deterministic Markov Process, where patients alternate between remission and relapse phases, and control the model via long-term cost function optimization. This considers treatment side effects, visit burden, and quality of life, using noisy blood marker measurements for feedback. Instead of discretizing the problem with a discrete Markov Decision Process, we apply the Partially-Observed Monte-Carlo Planning algorithm to solve the continuous-time, continuous-state problem, leveraging the near-deterministic nature of cancer progression. Our approach, tested on multiple myeloma patient data, outperforms exact solutions of the discrete model and allows greater flexibility in cost function modeling, enabling patient-specific follow-ups. This method can also be adapted to other diseases.
设计针对患者个体的随访策略是个性化癌症治疗的关键。能够基于患者偏好和医疗数据协助医生进行治疗决策并安排随访的工具将非常有益。这些工具应纳入治疗下疾病进展的现实模型、治疗策略的多目标优化，以及通过考虑患者病史来实现随访个性化的高效算法。我们建议使用分段确定性马尔可夫过程对癌症演变进行建模，患者在缓解期和复发期之间交替，并通过长期成本函数优化来控制该模型。这考虑了治疗副作用、就诊负担和生活质量，并使用有噪声的血液标志物测量值进行反馈。我们不是用离散马尔可夫决策过程对问题进行离散化，而是应用部分观测蒙特卡罗规划算法来解决连续时间、连续状态问题，利用癌症进展的近似确定性性质。我们的方法在多发性骨髓瘤患者数据上进行了测试，优于离散模型的精确解，并在成本函数建模方面具有更大的灵活性，从而实现针对患者个体的随访。该方法也可适用于其他疾病。
