Digital Biomarkers in Parkinson's Disease: Missing the Forest for the Trees?

Recently proposed digital measures of Parkinson's disease are gaining momentum. ¹ , ² More objective, more precise, and more easily repeated than human expert‐based assessments such as the current gold standard Movement Disorder Society Unified Parkinson's Disease Rating Scale (MDS‐UPDRS), ³ these measures—often conceptualized as digital biomarkers—promise a new dawn in personalized medicine. Digital tools are already being used as surrogate markers in clinical trials ⁴ and are advocated to improve clinical decision‐making. ⁵ Therefore, is it only a matter of more time, more data, and of course more money before at‐home digital assessments become the clinical gold‐standard? In this viewpoint we step‐back and re‐evaluate what it is we are trying to measure and why we are measuring it. From this perspective we argue that the current pursuit of more precise and objective measures may, if unchecked, lead to ever increasing bias in clinical trials and ever decreasing use for individual decision‐making. We may improve the statistical power of clinical trials, but at the cost of personalization. We propose a more holistic digital approach as a solution.

The Single Digital Biomarker Hypothesis

Many current digital measures aim to represent the severity of Parkinson's disease in an individual with a single number ⁶ , ⁷ , ⁸ although there are notable exceptions. ⁴ , ⁹ , ¹⁰ These proposed digital biomarkers are often more sensitive to clinical change than human expert ratings such as the MDS‐UPDRS, ¹¹ but the clinical interpretation and use of these scores remains unclear. Let us define a hypothetical smartphone‐derived digital biomarker—a single overall severity score that ranges from 0 to 10 where higher scores represent greater severity. If an intervention reduces an individual's score from 8.5 to 8.1, what does this mean? Often, this is interpreted as an improvement in the clinical features of disease and by proxy, an improvement in the underlying neuronal dysfunction whether modifying individual disease‐related pathophysiological processes or not. We call this interpretation the single digital biomarker hypothesis (Fig 1A). Much current regulatory, academic, and commercial thought ¹ often assumes this hypothesis to be true and is focused on overcoming the technical and practical barriers to validating these measures as digital endpoints. Here, we challenge the hypothesis itself and consider the reasons why improvements in overall digital measures do not necessarily imply overall improvements in clinical or neuronal dysfunction. We outline some of the conceptual credentials a digital measure must have if it is to be accepted as a digital biomarker. The foundation of our challenge stems from the well‐established view of Parkinson's disease as a multi‐dimensional, multi‐etiology syndrome.

FIG 1.

(A) The single digital biomarker hypothesis assumes that overall digital measures reflect clinical and neural dysfunction. Given a hypothetical digital measure of disease (higher is worse), a reduction in a digital measure (↓) can be interpreted as clinical improvement with associated reduction of neural dysfunction (↓). (B) A broader conceptualization of Parkinson's disease as a multi‐domain syndrome, however, provides two alternative reasons for this reduction in digital measure. Current measures incorporate only a few of the many clinical features associated with Parkinson's disease. In this example, the digital measure incorporates bradykinesia and tremor, but does not incorporate other clinical features such as anxiety, orthostatic hypotension, and drowsiness (only a few out of many domains are shown to aid legibility, others are represented with an ellipsis). Because interventions unevenly affect different clinical domains, a reduced digital measure (↓) because of improved measured bradykinesia may hide coincident meaningful worsening (↑) of unmeasured anxiety, orthostatic hypotension, and drowsiness. The digital measure reports an overall improvement, but the overall condition is in reality worse and the underlying neuronal dysfunction worse or the same (Δ). The digital measure is insufficient. (C) Similarly, unmeasured confounders can affect motor‐centric digital measures. Treated osteoarthritis of the knees may improve gait speed and subsequently reduce a digital measure incorporating it without any associated changes to the underlying Parkinson's disease. The arrows between the nodes of the graphs above represent directed causal links. Red labels and arrows represent the flow of information entering the digital measurement, whereas black labels and arrows represent information that the measure fails to incorporate. “Neuronal dysfunction” here is a broad term that does not necessarily refer to the individualized molecular/biological etiology.

Parkinson's Disease Cannot Be Measured By a Single Number

There is no unitary biological marker for Parkinson's disease progression—no equivalent of HbA1c for diabetes. ¹ We can measure clinical change, but this is difficult. Individuals with Parkinson's disease suffer in different ways—some may have noticeable hand tremor, others difficulty walking and yet others may be overwhelmed by anxiety. Individual severity, progression, and treatment response, therefore, occur unevenly across multiple clinical dimensions, likely reflecting multiple underlying biological and environmental factors. The multi‐system nature of Parkinson's disease was established early on and is to some extent reflected in current outcome scales such as the MDS‐UPDRS, ³ the Parkinson's Disease Quality of Life Questionnaire (PDQ‐39) ¹² and the Non‐Motor Symptoms Scale. ¹³ However, the vast majority of current digital measures focus only on measuring a limited aspect of Parkinson's disease. Wrist‐worn wearables are limited to motor observations at a single joint, ⁵ , ¹⁴ , ¹⁵ whereas smartphone assessments are dominated by motor assessments of one or a few limbs, sometimes with voice. ⁶ Exceptionally, a wider motor phenotype or even a brief cognitive test is included, ⁷ but these still fall drastically short of a holistic assessment of Parkinson's disease. Some reports describe their digital biomarker as a motor score rather than an overall score, but the same issue applies—typically only limited motor features are measured in limited parts of the body offering a limited view of the motor state. How does this issue affect the interpretation of digital measures as biomarkers? At the individual level, it affects our ability to take action solely on the basis of the “hyper‐precise” digital measure. It is clear that an intervention that improves our smartphone‐derived hypothetical digital score from 8.5 to 8.1 could easily worsen an individual's overall condition—measured clinical features such as bradykinesia may have improved slightly, but unmeasured features such as anxiety or dyskinesias may have deteriorated more substantially (Fig. 1B). Even with a digital measure, we still need a granular clinical assessment before we can recommend a change to an individual patient's treatment schedule. A selective digital measure will also impact group level inference, for example, in interventional clinical trials. An overall digital endpoint that is dominated by upper limb bradykinesia, for example, will be sensitive to interventions that improve hand motor function but relatively insensitive to interventions that improve other aspects of the condition. In the long‐term, unless we are careful, digital endpoints may bias translational science in favor of a particular mono‐ or oligosymptomatic profile.

Digital Measures Can Be Confounded

A second problem of the single digital biomarker hypothesis is the vulnerability of motor‐centric digital measures to confounding variables (Fig. 1C) and other sources of variation. ¹⁶ Here, systematic changes in a digital measurement are caused by unmeasured confounds rather than the neural dysfunction of Parkinson's disease. Unlike expert‐rated assessments such as the MDS‐UPDRS, digital measures are typically taken at‐home without supervision. Rhythmic hand movements while washing dishes or playing a musical instrument may be falsely interpreted as tremor or dyskinesia by an at‐home‐wearable. ¹⁴ Walking carefully on an icy pavement could be misinterpreted as bradykinesia. Many aspects of human life unrelated to the neural dysfunction of Parkinson's disease will interact in unknown ways with digital measurement. ¹⁶ Active task‐based measures—where the patient is performing a particular task such as tapping a smartphone screen —provide some level of control against unmeasured environmental confounds, but are still susceptible to another kind. Aging and co‐morbid diseases such as osteoarthritis will affect movement and confound digital measurements. Going back to our smartphone‐derived hypothetical digital measure, a drop from 8.5 to 8.1 could be an improvement in their Parkinson's disease, but may also represent cessation of their trombone‐playing practice or improved control of joint‐pain in their hands. At the group‐level within a randomized controlled trial, one may expect such confounds to “balance out” in the long run, but they would certainly add noise to the digital measure and reduce its sensitivity to change.

What Would a Digital Biomarker Look Like?

We have described how single digital measurements although precise often provide an insufficient or confounded measure of Parkinson's disease progression. They cannot be used as the sole basis of personalized clinical decision‐making. They show promise ⁴ but risk bias if used without care in clinical trials. How can we address this? A first step would be to develop digital measures that are holistic rather than reductive, measuring multiple clinical dimensions with granular detail—similarly to how a clinician would assess a patient. Note that these multiple measures must capture broad aspects of the patient's condition, not just be multiple measures of the same aspect (eg, hand tremor). With more faithful capture of a patient's condition, it will be clearer if symptoms have meaningfully worsened in one domain even if they have improved in others. If scores highlight a meaningful worsening of anxiety after an intervention while walking has improved, then it may be rational to withdraw the intervention even if our hypothetical single digital biomarker suggests an overall improvement. Unlike a single digital biomarker, therefore, a comprehensive digital assessment has the ability to assist clinical decision‐making at the individual level. Likewise, some types of confounded measurement would be more easily brought to light. Consider an intervention for osteoarthritis that improves walking speed by improving joint pain and stability. A single digital biomarker that incorporates a gait assessment may register improvement leading to the false conclusion that the individual's Parkinson's disease has got better. With multiple measures, however, it may be clearer that walking has improved because of a separate reason (in line with pain reduction and perhaps without associated improvement of hand function or tremor).

Although necessary, we do not claim here that increasing the breadth of assessment is sufficient for the development of an actionable multi‐domain measure—digital or otherwise. The PDQ‐39 ¹² is broad, but lacks the granularity required for individual clinical decision‐making. The MDS‐UPDRS ³ is also broad, but perhaps too blunt and unwieldy for routine clinical use (except as part of the assessment for advanced therapies). There are many other questions that novel digital measures will have to also address. We must understand what features are meaningful for patients, ¹⁷ be concrete in exactly what is being measured, ¹ and make sure digital tools are safe, robust and reliable. ² , ¹⁶ , ¹⁸ We must be aware that some aspects such as the number of falls in the last year are countable and easily measured, whereas other aspects such as the loss of confidence following a fall are not and so harder to quantify. Perhaps many important aspects are immeasurable.

We suggest beginning with a multi‐domain digital diary, perhaps focusing on interpretable countable observations (number of falls, episodes of urinary incontinence, hours slept in the daytime, etc) where possible. With appropriate care, this coarse, but curated and actionable resource, could provide the data required to develop more complex and interactive digital tools that incorporate objective measures. A similar development pipeline has already been proposed for the monitoring of clinical fluctuations in Parkinson's disease. ¹⁹ This may seem a backward step to developers of digital health technologies who seek to “upgrade the practice of medicine to one that is high‐definition”. ²⁰ They seek to reduce human inter‐rater error and increase objective measurement precision to provide better clinical endpoints for trials. However, in the pursuit of a more precise aggregate progression score, the individual has slipped out of focus. Comprehensiveness more so than precise, but unidimensional measurement is required for actionable, personalized care decisions. This will be familiar, even mundane, to expert clinicians who perform individualized multi‐dimensional assessments routinely as part of their clinical practice. However, somewhere along the digital development pathway this wisdom has been lost in translation.

Author Roles

(1) Research project: A. Conception, B. Organization, C. Execution; (2) Statistical Analysis: A. Design, B. Execution, C. Review and Critique; (3) Manuscript Preparation: A. Writing of the First Draft, B. Review and Critique.

A.J.: 1A, 1C, 3A, 3B.

A.J.E.: 1C, 3B.

A.J.L.: 1C, 3B.

Disclosures

Ethical Compliance Statement: The authors confirm that the approval of an institutional review board was not required for this work and informed consent was not necessary. We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this work is consistent with those guidelines.

Funding Sources and Conflicts of Interest: A.J. has been involved in the development and clinical assessment of a smartphone‐based tool for Parkinson's disease (cloudUPDRS). A.J. is funded by the Wellcome Trust (213038) and supported by the National Institute of Health Research University College London Hospitals Biomedical Research Centre.

Financial Disclosures for the Previous 12 Months: A.J. has received travel expenses from Merz to attend a scientific course. A.J.E. has received grant support from the National Institutes of Health and The Michael J. Fox Foundation; personal compensation as a consultant/scientific advisory board member for Neuroderm, Amneal, Acadia, Acorda, Bexion, Kyowa Kirin, Sunovion, Supernus (formerly, USWorldMeds), Avion Pharmaceuticals, and Herantis Pharma; personal compensation as honoraria for speakership for Avion and Amneal; and publishing royalties from Lippincott Williams and Wilkins, Cambridge University Press, and Springer. He cofounded REGAIN Therapeutics (a biotech start‐up developing nonaggregating peptide analogues as replacement therapies for neurodegenerative diseases) and is co‐owner of a patent that covers synthetic soluble nonaggregating peptide analogues as replacement treatments in proteinopathies.