Blood glycemic control is crucial for minimizing severe side effects in diabetes mellitus. Currently, two opposing treatment approaches exist: in formulaic methods, insulin care is calculated by parameter-based computation (i.e., correction factor, insulin-to-carb ratio, and absorption duration), which are fixed by the medical team based on the history of a tested patient blood glucose levels (BGLs). Alternatively, closed-loop methods test glycemic level via sensors and provide insulin boluses based on sensor data thus ignoring other medical information. Unlike the body, both these systems are reactive – chasing insulin dosage based on fluctuating BGL – resulting in significant fluctuations of glucose values, rather than the relatively flat profile normal to the body's glycemic control. Extended periods of these fluctuations – particularly high BGLs (hyperglycemia) result in vascular and organ epithelial damage, which increases comorbidities and is ultimately life-threatening. We propose an individualized treatment scheme based on machine learning artificial intelligence, which combines the best of both approaches and is tailored to the individual. We model patient reaction to insulin treatment as Markov decision process (MDP) thus allowing the system to find a unique, individualized and dynamically updating insulin care policy that would lead to flat blood glucose profiles in target areas. We incorporate an individualized “health reward function”, preferably from the medical team, describing a grading scheme of BGL tailored to the patient for even more precise glycemic control. The solution to MDP is found via reinforcement learning, which yields an individualized, optimal insulin care policy. This policy can prevent hypoglycemia, minimize high glucose duration and glycemic fluctuations. It can be further updated as the patient undergoes environmental changes. Significantly, our method provides the care team a constantly updated patient model, allowing them to better understand and support the patient.

血糖控制对于减少糖尿病的严重副作用至关重要。目前，存在两种对立的治疗方法：在公式化方法中，胰岛素护理是通过基于参数的计算（即校正因子、胰岛素与碳水化合物的比率和吸收持续时间）来计算的，这些由医疗团队根据病史确定被测试的患者血糖水平(BGLs)。或者，闭环方法通过传感器测试血糖水平，并根据传感器数据提供胰岛素丸剂，从而忽略其他医疗信息。与身体不同，这两个系统都是反应性的——根据波动的 BGL 追踪胰岛素剂量——导致葡萄糖值的显着波动，而不是身体血糖控制正常的相对平坦的曲线。这些波动的延长时间——特别是高 BGL（高血糖）会导致血管和器官上皮损伤，从而增加合并症并最终危及生命。我们提出了一种基于机器学习人工智能的个性化治疗方案，它结合了两种方法的优点，并为个人量身定制。我们将患者对胰岛素治疗的反应建模为马尔可夫决策过程 (MDP)，从而使系统能够找到独特的、个性化的和动态更新的胰岛素护理策略，从而使目标区域的血糖曲线保持平稳。我们结合了个性化的“健康奖励功能”，最好是来自医疗团队，描述了为患者量身定制的 BGL 分级方案，以实现更精确的血糖控制。MDP 的解决方案是通过强化学习找到的，这会产生个性化的最佳胰岛素护理策略。该政策可以防止低血糖，最大限度地减少高糖持续时间和血糖波动。当患者经历环境变化时，它可以进一步更新。重要的是，我们的方法为护理团队提供了一个不断更新的患者模型，使他们能够更好地了解和支持患者。