Predicting motor, cognitive & functional impairment in Parkinson's

Abstract

Objective: We recently demonstrated that 998 features derived from a simple 7-minute smartphone test could distinguish between controls, people with Parkinson's and people with idiopathic Rapid Eye Movement sleep behavior disorder, with mean sensitivity/specificity values of 84.6-91.9%. Here, we investigate whether the same smartphone features can be used to predict future clinically relevant outcomes in early Parkinson's.

Methods: A total of 237 participants with Parkinson's (mean (SD) disease duration 3.5 (2.2) years) in the Oxford Discovery cohort performed smartphone tests in clinic and at home. Each test assessed voice, balance, gait, reaction time, dexterity, rest, and postural tremor. In addition, standard motor, cognitive and functional assessments and questionnaires were administered in clinic. Machine learning algorithms were trained to predict the onset of clinical outcomes provided at the next 18-month follow-up visit using baseline smartphone recordings alone. The accuracy of model predictions was assessed using 10-fold and subject-wise cross validation schemes.

Results: Baseline smartphone tests predicted the new onset of falls, freezing, postural instability, cognitive impairment, and functional impairment at 18 months. For all outcome predictions AUC values were greater than 0.90 for 10-fold cross validation using all smartphone features. Using only the 30 most salient features, AUC values greater than 0.75 were obtained.

Interpretation: We demonstrate the ability to predict key future clinical outcomes using a simple smartphone test. This work has the potential to introduce individualized predictions to routine care, helping to target interventions to those most likely to benefit, with the aim of improving their outcome.

Figure 1

Smartphone models. In the search for a scalable solution to the quantification of motor symptoms in Parkinson's, an Android based smartphone app was installed on a range of consumer grade smartphones that were used in clinic and provided to participants to take home. Participants also had the option of being sent a link to download the app onto their own Android smartphone. A specialized smartphone app was used to collect the raw accelerometer, microphone and screen data and was run alongside KitKat, Lollipop, Marshmallow, Nougat, and Oreo Android operating systems. The raw data from the app was encrypted, time‐stamped, and uploaded to a secure online server. The processing and analysis of the data was performed separately using computer‐based Matlab® software (R2018a; Mathworks®, USA). “Others” include: Samsung Galaxy Ace 4 SM‐G357FZ, Samsung Galaxy Ace 2 GT‐I8160, Samsung Galaxy S3 Mini GT‐I8200N, LG Optimus 3G CX670, Samsung Galaxy S III mini I8190, Samsung Galaxy J5 J500FN, Sony Xperia L C2105, Moto G LTE XT1039, Huawei Ascend G510, Samsung Galaxy S4 I9505.

Figure 2

Flow charts demonstrating the time windows whose recordings were included in the analyses of the future onset of (A) falls, (B) freezing, (C) postural instability, (D) cognitive impairment, (E) difficulty doing hobbies, and (F) need for help at home.

Figure 3

Receiver operating characteristic curves for classification by random forests in the prediction of the future onset of (A) falling, (B) freezing, (C) postural instability, (D) cognitive impairment, (E) difficulty doing hobbies and (F) the need for help. The diagonal dotted line corresponds to an AUC of 0.50 and indicates an uninformative model. The false positive rate (1‐specificity) is shown on the x axis and the true positive rate (sensitivity) is shown on the y axis.