Multi-Hour Blood Glucose Prediction in Type 1 Diabetes: A Patient-Specific Approach Using Shallow Neural Network Models
Multi-Hour Blood Glucose Prediction in Type 1 Diabetes: A Patient-Specific Approach Using Shallow Neural Network Models
复制标题
DOI：
10.1089/dia.2020.0061
复制
发表时间：
2020-10-13
影响因子：
5.4
通讯作者：
Sankaranarayanan, Sriram
中科院分区：
文献类型：
--
作者：
Kushner, Taisa;Breton, Marc D.;Sankaranarayanan, Sriram
背景：考虑到目前的胰岛素作用特征和对胰岛素的血糖反应的性质，迫切需要更长期、准确的血糖预测，以告知胰岛素给药方案，并为1型糖尿病（T1 D）的治疗提供有效的决策支持。然而，目前的方法只能达到可接受的准确度预测范围为1小时，而典型的餐后波动和胰岛素作用曲线持续4-6小时。在这项研究中，我们提出了利用“浅层”神经网络开发的60-240分钟预测范围的模型，与相关方法相比，其复杂性显着降低。方法：开发基于患者特异性神经网络的预测模型，并对先前从24例T1 D受试者队列中收集的数据进行测试。模型的设计是为了避免严重的陷阱，通过将基本的生理知识纳入模型结构。生成患者特定模型以预测提前60、90、120、180和240分钟的葡萄糖，以及“转移学习”方法以提高数据有限的患者的准确性。最后，我们确定了导致整体模型准确度更高的子组特征。结果：60、90、120、180和240 min的均方根误差分别为28 +/- 4、33 +/- 4、38 +/- 6、40 +/- 8和43 +/- 12 mg/dL。对于所有的预测范围，至少93%的预测是临床上可接受的克拉克误差网格。历史连续葡萄糖监测（CGM）值的方差是迁移学习方法需求的一个强有力的预测因素。结论：使用从过去CGM数据和胰岛素日志中提取的特征的浅层神经网络可以以令人满意的准确度实现多小时葡萄糖预测。模型是患者特异性的，在不需要额外测试的情况下根据现成的数据进行学习，并且与相关方法相比提高了准确性，同时降低了复杂性，为能够涵盖大部分胰岛素作用时间框架的新咨询和闭环算法铺平了道路。
Background:Considering current insulin action profiles and the nature of glycemic responses to insulin, there is an acute need for longer term, accurate, blood glucose predictions to inform insulin dosing schedules and enable effective decision support for the treatment of type 1 diabetes (T1D). However, current methods achieve acceptable accuracy only for prediction horizons of up to 1 h, whereas typical postprandial excursions and insulin action profiles last 4-6 h. In this study, we present models for prediction horizons of 60-240 min developed by leveraging "shallow"neural networks, allowing for significantly lower complexity compared with related approaches. Methods:Patient-specific neural network-based predictive models are developed and tested on previously collected data from a cohort of 24 subjects with T1D. Models are designed to avoid serious pitfalls through incorporating essential physiological knowledge into model structure. Patient-specific models were generated to predict glucose 60, 90, 120, 180, and 240 min ahead, and a "transfer learning" approach to improve accuracy for patients where data are limited. Finally, we determined subgroup characteristics that result in higher model accuracy overall. Results:Root mean squared error was 28 +/- 4, 33 +/- 4, 38 +/- 6, 40 +/- 8, and 43 +/- 12 mg/dL for 60, 90, 120, 180, and 240 min, respectively. For all prediction horizons, at least 93% of predictions were clinically acceptable by the Clarke error grid. Variance of historic continuous glucose monitor (CGM) values was a strong predictor for the need of transfer learning approaches. Conclusions:A shallow neural network, using features extracted from past CGM data and insulin logs, can achieve multi-hour glucose predictions with satisfactory accuracy. Models are patient specific, learnt on readily available data without the need for additional tests, and improve accuracy while lowering complexity compared with related approaches, paving the way for new advisory and closed loop algorithms able to encompass most of the insulin action timeframe.