Wearables-derived risk score for unintrusive detection of α-synuclein aggregation or dopaminergic deficit

Abstract

Background: Smartwatch data has been found to identify Parkinson's disease (PD) several years before the clinical diagnosis. However, it has not been assessed against the gold standard but costly and invasive biological and pathological markers for PD. These include dopaminergic imaging (DaTscan) and cerebrospinal fluid alpha-synuclein seed amplification assay (SAA), which are being studied as markers thought to represent the onset of PD pathology.

Methods: Here, we combined clinical and biological data from the Parkinson's Progression Marker Initiative (PPMI) cohort with long-term (mean: 485 days) at-home digital monitoring data collected using the Verily Study Watch. We derived a digital risk score based on sleep, vital signs, and physical activity features to distinguish between PD (N = 143) and healthy controls (N = 34), achieving an area under precision-recall curve of 0.96 ± 0.01. We compared it with the Movement Disorder Society (MDS) research criteria for prodromal PD to detect dopaminergic deficit or α-synuclein aggregation in an at-risk cohort consisting of people with genetic markers or prodromal symptoms without a diagnosis of PD (N = 109, mean age = 64.62 ± 6.86, 40 men and 69 women).

Findings: The digital risk correlated with the MDS research criteria (r = 0.36, p-value = 1.46 × 10^-4) and was increased in individuals with subthreshold Parkinsonism (p-value = 4.99 × 10^-6) and hyposmia (p-value = 3.77 × 10^-2). The digital risk was correlated to a stronger degree with DaTscan putamen binding ratio (r = -0.32, p-value = 6.64 × 10^-4) than the MDS criteria (r = -0.19, p-value = 6.81 × 10^-3) but to a weaker degree with SAA (r = 0.2, p-value = 3.9 × 10^-2) than the MDS (r = 0.43, p-value = 1.3 × 10^-5). The digital risk score achieved higher sensitivity in identifying synucleinopathy or neurodegeneration (0.59) than the MDS score (0.35) but performed on-par with hyposmia (0.59) with a combination of hyposmia and digital risk score achieving the highest sensitivity (0.71). The digital risk score showed lower precision (0.18) than other models.

Interpretation: A digital risk score from smartwatch data should be further explored as a possible first sensitive screening tool for presence of α-synuclein aggregation or dopaminergic deficit followed by subsequent more specific tests to reduce false positives.

Funding: This project is funded by Welsh Government through Health and Care Research Wales, Medical Research Council (MRC), Higher Education Funding Council for Wales, UK Dementia Research Institute, Alzheimer's Society and Alzheimer's Research UK, Dementia Platforms UK, UKRI Engineering and Physical Sciences Research Council (EPSRC), NIHR Imperial Biomedical Research Centre (BRC), Great Ormond Street Hospital and the Royal Academy of Engineering, Edmond J. Safra Foundation, Ser Cymru II programme, and the European Regional Development Fund.

Keywords: Parkinson's disease; Prodromal; Risk modelling; Smartwatch.

Fig. 1

Digital measures capture differences in at-risk groups. The boxplots show the residual overall mean of digitally tracked sleep efficiency adjusted for age and sex with parameters learnt from a linear regression on the healthy controls. The overall mean is computed over the whole observation time per subject for each group. The boxplots depict the group median and quartiles per group with the whiskers showing the Q3 + 1.5 interquartile range (IQR) and Q1 − 1.5 IQR (Parkinson's disease cases: PD; healthy controls: HC; carriers of genetic risk alleles or prodromal symptoms without a diagnosis of PD: GBA, LRRK2, hyposmia, polysomnography-proven RBD, positive DaTscan, positive SAA; union of these: at-risk). The number in the yellow box indicates the number of individuals per group. Group differences were calculated with two-sided t-test comparing PD and HC to each of the at-risk groups. Lines and numbers show significant differences with 0.05 FDR corrected p-values.

Fig. 2

Derivation of risk scores and overview of statistical analyses. Overview of analysis. Derivation of risk scores and biological and pathological markers. Illustration of performed tests and modelling.

Fig. 3

Performance of digital risk models. The performances for the digital risk score models are shown compared to a) baseline, b) other machine learning models, c) other feature sets, and d) other considered time frames. The precision-recall curves are shown as the mean on the outer 5-folds of the nested cross-validation. The shaded area displays the 95% Confidence Interval (CI). For each classifier, the legend shows the mean area under the precision-recall curve (AUPRC) with the standard deviation. SVM: support vector machine, rbf: radial basis function, RF: random forest.

Fig. 4

Digital risk score correlates with MDS prodromal score and biological markers. The relation between the different risk scores and biological markers is shown. On the diagonal, the distribution for each diagnostic group is displayed (PD: diagnosed Parkinson's disease, HC: healthy control, Prodromal: at-risk cohort of genetic mutations carriers and individuals with prodromal symptoms). The scatterplot shows the relation between each pair of markers (digital: digital risk score, MDS restricted: Movement Disorder Society (MDS) prodromal risk score without DaTscan information, MDS: MDS prodromal risk score with DaTscan information if available, CSF α-synuclein SAA Fmax mean: mean value of the five repetitions of seed amplification assay (SAA) on CSF, DaTscan minimum putamen: minimum of hemispheres dopaminergic imaging scan (DaTscan) striatal binding ratio (SBR¹⁰) in putamen) in the at-risk group (N = 109) with the text box displaying the Pearson r coefficient and the 0.05 FDR corrected p-value.

Fig. 5

Digital risk score is increased in individuals with known prodromal markers and biological classification groups. a) The boxplots show the difference in digital risk score between carriers and non-carriers (x-axis). The 0.05 FDR-corrected p-value from two-sided Welch t-test is shown. The yellow box presents the number of subjects in each group. This plot shows those prodromal markers and risk factors from the model included in Heinzel, Berg that were significant after FDR-correction. A complete table with statistical results can be found in Supplementary Table S4. b) The distribution of risk scores for the different biological groups defined by SynNeurGe for the digital, the MDS, and the restricted MDS risk scores. c) The distribution of the risk score for the different biological stages defined by NSD for the digital, MDS, and the restricted MDS risk scores.