Turner Dana P, Caplis Emily, Patel Twinkle, Houle Timothy T
Department of Anesthesia, Critical Care and Pain Medicine, Massachusetts General Hospital, Harvard Medical School, Boston, MA, USA.
medRxiv. 2025 May 5:2025.05.03.25326924. doi: 10.1101/2025.05.03.25326924.
To extend the application of surprisal theory for predicting migraine attack risk by developing methods to estimate trigger variable likelihood in real time, under conditions of limited personal observation.
Prior work has demonstrated that higher surprisal, a measure quantifying the unexpectedness of a trigger exposure, predicts headache onset over 12 to 24 hours. However, these analyses relied on retrospective expectations of trigger exposure formed after extended data collection. To operationalize surprisal prospectively, Bayesian methods could update expectations dynamically over time.
In a prospective daily diary study of individuals with migraine (N = 104), data were collected over 28 days, including stress, sleep, and exercise exposures. Bayesian models were applied to estimate daily expectations for each variable under uninformative and empirical priors derived from the sample. Stress was modeled using a hurdle-Gamma distribution, sleep using a rounded Normal distribution, and exercise using a Bernoulli distribution. Surprisal was calculated based on the predictive distribution at each time point and compared to static empirical surprisal values obtained after full data collection.
Dynamic Bayesian surprisal values systematically differed from retrospective empirical estimates, particularly early in the observation period. Divergence was larger and more variable under uninformative priors but attenuated over time. Empirically informed priors produced more stable, lower-bias surprisal trajectories. Substantial individual variability was observed across exposure types, especially for exercise behavior.
Prospective surprisal modeling is feasible but highly sensitive to prior specification, especially in sparse data contexts (e.g., a binary exposure). Incorporating empirical or individually informed priors may improve early model calibration, though individual learning remains essential. These methods offer a foundation for real-time headache forecasting and dynamic modeling of brain-environment interactions.
通过开发在个人观察有限的情况下实时估计触发变量可能性的方法，扩展惊奇理论在预测偏头痛发作风险方面的应用。
先前的研究表明，较高的惊奇值（一种量化触发暴露意外程度的指标）可预测12至24小时内的头痛发作。然而，这些分析依赖于在长时间数据收集后形成的触发暴露的回顾性预期。为了前瞻性地实施惊奇值，贝叶斯方法可以随时间动态更新预期。
在一项针对偏头痛患者（N = 104）的前瞻性每日日记研究中，收集了28天的数据，包括压力、睡眠和运动暴露情况。应用贝叶斯模型在从样本得出的无信息先验和经验先验下估计每个变量的每日预期。压力使用障碍-伽马分布建模，睡眠使用舍入正态分布建模，运动使用伯努利分布建模。根据每个时间点的预测分布计算惊奇值，并与完整数据收集后获得的静态经验惊奇值进行比较。
动态贝叶斯惊奇值与回顾性经验估计值系统地不同，特别是在观察期早期。在无信息先验下，差异更大且更具变异性，但随着时间推移会减弱。基于经验的先验产生更稳定、偏差更小的惊奇值轨迹。在不同暴露类型中观察到了显著的个体差异，尤其是运动行为方面。
前瞻性惊奇值建模是可行的，但对先验设定高度敏感，特别是在稀疏数据情况下（例如二元暴露）。纳入经验或个体特定的先验可能会改善早期模型校准，不过个体学习仍然至关重要。这些方法为实时头痛预测和脑-环境相互作用的动态建模提供了基础。
