Predicting Parkinson's disease from wrist worn accelerometer data from UK Biobank

帕金森病（PD）是全球第二大最常见的神经退行性疾病，2020年的患病率约为940万。该疾病是由黑质致密部多巴胺能神经元逐渐退化引起的，导致运动能力改变相关症状。虽然这种情况通常出现在生命的后期阶段，但这种疾病的患病率继续上升，从2016年到2020年增加了57%。据预测，到2040年，神经退行性疾病将超过癌症，成为疾病相关死亡的主要原因。虽然PD不会直接导致死亡，但它会显著降低患者的生活质量，并且有许多PD相关症状可能是致命的。帕金森氏症的一个特别问题是，目前没有治愈方法，也没有已知的明确诊断方法。目前的工作是探索使用英国生物银行的数据来建立一个高斯混合模型（GMM），从380名总受试者中区分202名PD受试者，给定一周的加速度计，从腕部佩戴的智能手表[1]中提取。这种做法取得了合理的成功；然而，也注意到一些局限性，这可能会导致在这一领域的进一步研究。本文使用的GMM采用了选定的数据驱动特征，因此没有利用已知的运动相关PD症状，包括震颤、高频震动、运动迟缓、运动缓慢，也没有利用能够提取这些特征的一系列模型。此外，缺乏对模型假阳性的分析，这可能会给出与PD患者相似的行为指示，从而更好地理解预测。目的：建立PD检测的深度学习模型：研究将从使用变压器网络从英国生物银行数据集的步态周期预测PD开始。步态周期将首先从跨越1周的完整加速度计中提取。这将使用用于从加速度计识别步态的已知数据集进行验证。然后，这些步态周期将被输入到变压器网络，以提供PD检测器。制定PD相关运动症状监测的严重程度评分：对于监测目的而言，能够识别对PD相关运动症状有积极或消极影响的干预措施是有用的。要做到这一点，有必要有一个分数，给出PD相关运动系统的严重程度和规律性的指示，给出在给定时间窗内观察到的PD的严重程度的指示。要做到这一点，可以使用与用于诊断预测PD的模型的概率输出相关的分数。了解有助于帕金森病预测的相关特征：由于帕金森病没有明确的诊断，因此从用于预测帕金森病的算法中收集尽可能多的知识是有用的。从加速度计数据中提取对模型操作至关重要的特征是可能的。然后可以将这些特征与专家/已知文献进行比较，以建立更好的PD诊断共识。这可以通过使用各种可解释的AI技术来实现。对于类似的运动相关疾病，应用实现上述目标时使用的原则：有许多病症和疾病可以通过运动能力的变化来指示，并且可以通过加速度计来拾取特征。这包括类风湿关节炎、肌萎缩性侧索硬化症、肯尼迪氏病等。在这项研究中使用的技术框架也有可能应用于这些疾病的诊断和监测。这项工作是与葛兰素史克公司合作进行的。该项目属于EPSRC医疗保健技术主题，与临床技术研究领域相关。参考文献：b[1] Williamson等人（2021）从英国生物银行传感器的腕带加速度计检测帕金森病，21(6),2047

Parkinson's disease (PD) is the second most common neurodegenerative disease worldwide, with a prevalence of roughly 9.4M in 2020. The disease is caused by the gradual degeneration of dopaminergic neurons in the substantia nigra pars compacta, which leads to symptoms related to changes in motor ability. Though the condition often presents during the later stages of life, the prevalence of this disease continues to rise, with a 57% increase from 2016 to 2020. It is predicted that by 2040, neurodegenerative diseases will surpass cancer as a leading cause of disease related death. While PD does not directly cause death, it can significantly reduce the quality of life of sufferers and there are many PD related symptoms that can be fatal. A particular concern with PD, is that it currently has no cure, and no known definitive test for diagnosis. Current work is exploring the use of UK Biobank data to build a Gaussian Mixture Model (GMM) that distinguishes 202 PD subjects from 380 total subjects, given a week of accelerometery extracted from wrist worn smart watches [1]. This found reasonable success; however, several limitations were noted, which give rise to potential further research in this field. The paper uses a GMM that takes chosen data-driven features, hence does not take advantages of known motor related PD symptoms, including tremors, a high frequency shaking, and bradykinesia, a slowness of movement, nor a range of models capable of extracting these features. In addition, there is a lack of analysis of false positives of the model, which may give an indication of behaviours that resemble that of PD sufferers, and therefore a better understanding of the prediction. Aims: Build a deep learning model for PD detection: The investigation will begin with the use of transformer networks to predict PD from periods of gait from the UK biobank data set. The periods of gait will first be extracted from the full accelerometery spanning 1 week. This will be validated using known datasets used for identifying gait from accelerometery. These gait periods will then be input to a transformer network to deliver a PD detector. Develop a severity score for monitoring PD related motor symptoms: It is useful for purposes of monitoring to be able to identify interventions that have either positive or negative effects on PD related motor symptoms. To do so, it is necessary to have a score that gives an indication of the severity and regularity of PD related motor systems, to give an indication of the severity of PD observed over a given time window. To do so, it possible to use a score related to the probability output by the model used to diagnostically predict PD. Develop an understanding of relevant features that aid PD prediction: As there is no definitive diagnosis for PD, it is useful to gather as much knowledge as possible from algorithms used to predict PD. It may be possible to extract features from the accelerometer data taken that are critical in the model operation. These features can then be compared to experts/known literature, to build a better consensus for PD diagnosis. This is possible using various explainable AI techniques. Apply the principles used in achieving the previous aims, for similar motor related diseases: There are many conditions and diseases that may be indicated by a change in motor ability, and features that can be picked up by accelerometery. This includes Rheumatoid Arthritis, Amyotrophic Lateral Sclerosis, Kennedy's disease etc. It is possible that the framework for techniques used during this research can also be applied to the diagnosis and monitoring of those diseases. This work is in collaboration with GlaxoSmithKline plc. This project falls within the EPSRC Healthcare technologies Theme and is related to the Clinical technologies research area. Reference: [1] Williamson et al (2021) Detecting Parkinson's Disease from Wrist-Worn Accelerometry in the UK Biobank Sensors, 21(6), 2047