The myth of digital biomarkers in Alzheimer's disease: how to make them a reality

Abstract

Purpose of review

With an estimated 41.1B digital devices, the term “digital biomarkers” has been increasingly bandied about in the research literature. There is, however, a significant disconnect between the presumption of digital biomarkers and the reality of digital biomarkers.

Recent findings

The research literature embraces the concept of digital biomarkers without concomitant evidence for validation of digital measures as biomarkers. Unlike imaging or blood-based biomarkers, there is a woeful lack of research dedicated to validating digital measures as biomarkers. This gap also presents an opportunity. Regulatory agencies worldwide have long-standing protocols used by pharmaceutical and biotech companies to stand up quality management systems (QMS) that track research from inception to regulatory approved submissions. The recent United States (US) Food and Drug Administration (FDA) approval of Alzheimer's disease (AD) plasma biomarkers is another example where successful QMS implementation provided the processes and transparency necessary to obtain approval. Regulatory guidelines for digital technology validation are more circumspect on validation pathways of AD digital biomarkers, but FDA provides a framework for building a QMS that could potentially do so.

Summary

Building an open source QMS for AD digital biomarker validation will be a critical breakthrough for harnessing the potential of digital technologies for detection, monitoring and treatment of AD and related disorders.

INTRODUCTION

In the United States, the start of every calendar year is marked by two conferences, the JP Morgan Healthcare Conference held in San Francisco, California and the Consumer Electronic Show (CES) in Las Vegas, Nevada. Not infrequently, the dates of these two conferences overlapped because they served two different segments of the business world. But in less than a decade, CES has shifted from showcasing electronics for home use to include showcasing internet-of-things (IoT) connected devices that can monitor a myriad of physical and mental health behaviors. In 2019, there was a frenetic back-and-forth between the two conferences among those whose interest overlapped with both. Beginning in 2020, the dates of the two conferences no longer overlap. This adjustment in conference schedules is evidence of how much consumer level IoT technologies have penetrated into the healthcare arena. 

Box 1.

no caption available

CURRENT DIGITAL BIOMARKER LANDSCAPE REVIEW

According to the United States Food and Drug Administration (FDA), a biomarker is a “defined characteristic that is measured as an indicator of normal biological processes, pathogenic processes, or responses to an exposure or intervention, including therapeutic interventions.” They further specify that “a biomarker is not an assessment of how an individual feels, functions, or survives.” To assist researchers in the development and identification of biomarkers for effective translation from bench to bedside, the FDA-National Institutes of Health (FDA-NIH) Joint Leadership Council created a harmonized framework using the terminology “BEST” (Biomarkers, Endpoints, and other Tools [1]). The BEST glossary defines and distinguishes biomarkers from clinical assessments to ensure end products can effectively be translated into FDA-approved tools. Seven distinct categories of biomarkers range from diagnosis to prognosis to safety monitoring. In addition to the United States, the European Medicines Agency (EMA) has also been evaluating digital health technologies (DHT), informed by a scanning report [2,3]. But to date, the EMA has been focused on derived digital endpoints rather than defining a validation pathway for digital biomarkers. The United Kingdom (UK) Medicines and Healthcare products Regulatory Agency (MHRA) provides guidance for digital technologies that includes those using artificial intelligence (AI) via their Software and AI as a medical device pathway [4] as does Health Canada using their Software as a Medical Device (SaMD), Digital Health Technologies (DHTs), and Medical Device Regulations frameworks. Of note is that neither outlines specific guidelines for digital biomarkers. Australia's Therapeutic Goods Administration (TGA), China's National Medical Products Administration (NMPA), and Japan's Pharmaceutical and Medical Device agencies (PMDA) are similarly pursuing DHT's as clinical endpoints. Further, countries often look to the FDA, EMA and UK to inform their own regulatory policies [5^▪▪,6]. Currently, the FDA has the most developed guided framework to follow in determining what digital phenotypes qualify as digital biomarkers that regulatory agencies would approve. Given the current state globally, there remains a significant gap between the idea of digital biomarkers and clear provision of guidance from regulatory agencies on how to validate them.

THE REALITY OF DIGITAL BIOMARKERS

As of October 1, 2025, PubMed search of “digital biomarkers” produces a list of 8630 publications. As shown in Fig. 1, restricting the search to 2010–October 2025, accounts for 92.7% of these publications, and 66.7% fall between 2020 to October 2025. These numbers evidence that the concept of digital biomarkers has emerged relatively recently. The search terms “digital biomarkers” and “validation” and “FDA” result in a total of 31 publications, of which four involve IoT devices for collecting health related data [7^▪,8^▪,9,10]. The same search replacing FDA with EMA, produced 11 citations, one of which reported used of a mHealth application to aid in the management and care of rhinitis and asthma care [11]. Searches of “digital biomarkers” and “validation” and “China NMPA” and “Japan's Pharmaceutical and Medical Devices Agency (PMDA)” produced no relevant papers, while a Health Canada Agency search resulted in 9 citations. One of those papers included a dreaMS app that collected functional domains that were of significant utility to persons living with multiple sclerosis [12] while the second one represented a multi-consortium effort to standardized collection of data using wearable devices [13^▪▪]. None of these studies focused on specifically validating digital measures as biomarkers. This scraping of the research literature suggests that while the term “digital biomarker” has entered the accepted use lexicon, the actual science behind digital biomarkers is far more limited.

FIGURE 1.

Original figure illustrating the search terms “digital biomarker” on PubMud as of October 1, 2025, through four different date ranges. The last search includes the additional terms “validation” and “FDA”.

DIGITAL BIOMARKERS IN THE CONTEXT OF ALZHEIMER'S DISEASE

The U.S. FDA has led the way in the approval of in vivo AD biomarkers, dating back to the early 90s, through assays detecting amyloid and tau from cerebrospinal fluid (CSF), that initially received FDA clearance, but not FDA approval [14^▪▪,15]. In 2004 the discovery of Pittsburgh Compound B (PiB), a ligand that would bind to amyloid in the brain, led to visualization using positron emission tomography (PET) scans. The first PET amyloid test for AD was FDA approved in 2012 [16], followed by over a decade of additional research that resulted in the first FDA approval of a PET tau test in 2020 [17]. Most recently, FDA approved plasma-based biomarkers have emerged [18], marking a significant scientific advancement given their potential low-cost scalability. To achieve these FDA milestones, multiple studies were conducted spanning decades to accumulate the evidence that effectively demonstrated acceptable reliability and performance of these markers as specific to AD.

Figure 2 illustrates the results from a PubMed search strategy for “digital biomarkers and AD,” where 98.9% of publications are in the last decade (2015–October 1, 2025). Narrowing the search by adding the term “validation” reduced the number of publications to 120 (27.3%). Given that the FDA is the only regulatory agency that has specifically provided guidance on digital biomarkers, the term “FDA” was added into the search terms and resulted in a paucity of three publications [19^▪,20^▪▪,21^▪▪]. Only one of these publications related a measure of air pollutants collected using a digital instrument against those verified to be AD [20^▪▪]. Thus, a literature review for AD-specific digital biomarkers reveals the field remains wide open because of limited validation evidence.

FIGURE 2.

Original figure illustrating the funnel of all search results on PubMud for “digital biomarkers and AD” narrowing the scope to results within the past decade and then including the terms “validation” and then “FDA”.

Surprising is the juxtaposition between the amount of research conducted to achieve FDA approval of an AD biomarker via collection of CSF, imaging or blood and the lack of research to achieve FDA approval of an AD biomarker via collection through IoT devices.

Thus, while the idea of digital biomarkers is currently making waves within the research community, there is little understanding of how this translates into practice. With a projected 41.1 billion IoT-connected devices in 2025, there is immense opportunity to collect data required to identify and validate digital biomarkers across an array of health indices and do so in a minimally invasive, cost-conscious way at population scale. In the context of AD, cognitive impairment is considered one of the hallmark clinical indicators of disease. Cognitively related phenotypes such as those extracted from speech or typing behaviors on a smartphone, if properly validated, could represent digital biomarkers of clinically symptomatic AD [22–24,25^▪▪]. Using analysis of common human behavior collected through ubiquitous IoT devices to accurately extract AD clinical indices could upend current AD diagnostic practices and create greater care and treatment equity across geographic and socio-economic dimensions.

MOVING THE FIELD OF ALZHEIMER'S DISEASE DIGITAL BIOMARKERS FORWARD

A long-persistent challenge in AD is detecting early cognitive changes that are accurately reflective of an underlying progressive biological process [26]. Cognition is a notoriously variable behavior, where fluctuations can happen constantly from morning to night [27^▪▪]. Some people report better cognitive capacity when they first wake up, while others contend their greatest cognitive performance comes in the late-night hours. Physical activity during the day can increase cognitive skills in the hours that follow [28], and poor sleep at night can attenuate these same skills the next day [29,30^▪▪].

Legacy research methodologies are rapidly becoming obsolete in this emerging technology-driven landscape. Recent concerns about academic research reproducibility will only be further exacerbated as ubiquitous IoT devices get further integrated into research data collection protocols and fuel AI-powered computational approaches that lack transparency into how results are generated [31].

While NIH data sharing mandates have resulted in many repositories in which data from multiple studies are available, variability in data sharing practices is a significant contributor to the reproducibility quandary. Access to the same datasets is not sufficient because there are multiple subjective decisions made in preparing a dataset for analysis [32]. What are defined as outliers, how to impute missing data, whether to normalize a skewed distribution or convert to categorical variables, which co-variates to include are just a few examples of the decisions being made before analysis can commence. Preparing data into an analyzable format can be considered by some researchers as their intellectual property, leading to reluctance to share these prepared data. Further, there are no gold standards for these decisions, and thus, prepared datasets may not be adequate within context of use for other researchers. Reproducibility will only be possible if all researchers share all their prepared datasets, clear written descriptions of each data point included in the dataset, and all the programming code used to select, exclude, standardize and harmonize the data.

Artificial intelligence (AI)-powered analytics are compounding the reproducibility difficulty because of the lack of transparency on how deep learning and other machine learning systems arrive at any result. Further, AI model size has rapidly increased (billions and trillions of parameters for large language models (LLMs)) making them harder to share and computationally expensive to use (requiring access to high-end GPUs). In addition, foundational model training has shifted to industry for some models (e.g., LLMs), with models evolving fast and several prominent models not being open-sourced. Typically, prior model generations are also not maintained which impedes the reproducibility of earlier results. In use of data, bootstrapping and other resampling methods have helped to offset the lack of transparency, and regulatory agencies (e.g., FDA, EMA, MHRA, TGA, PMDA, etc.) do accept bootstrapping results as long as the data type and context of use are well defined, and the results are clinically meaningful.

DIGITAL VOICE AS AN ALZHEIMER'S DISEASE DIGITAL BIOMARKER STARTING POINT

Digital voice as a biomarker has some limited research support from studies validating speech and linguistic features with underlying neurobiological correlates. Acoustic features from audio recordings of traditional neuropsychological assessments were related to numerous neuroanatomical alterations as measured by brain MRI, including global volume loss, particularly within the medial temporal lobe [33^▪▪]. Growing evidence suggests digital voice is related to AD pathological biomarkers, where digital voice predicted amyloid status measured by cerebrospinal fluid or PET [34–37,38^▪▪] and tau status determined by PET scans [39] in asymptomatic individuals, with no concomitant associations found with traditional paper and pencil neuropsychological scores. In one of the few studies linking digital voice to plasma AD biomarkers, acoustic features of audio spectrum magnitude and a measure of acoustic timber were found to be related to tau and AB40, respectively [40].

Interest in digital voice as a potential biomarker will likely continue to grow. Speaking taps into multiple cognitive domains and universally most people speak. The deep penetration of mobile devices with recording capabilities such as smartphones has resulted in a globally feasible approach to collecting speech. The Alzheimer's Disease Research Centers Uniform Dataset (UDS) 4.0 newly launched protocol includes instructions on how to digitally record examiner administered neuropsychological testing. The Alzheimer's Drug Discovery Foundation is supporting a Speech Dx program that administers a standardized voice collection protocol paired with AD clinical and biomarker data from which to discover and validate voice features as potential AD voice biomarkers [41^▪▪]. Furthermore, many longitudinal epidemiological cohorts (i.e., Framingham Heart Study, Bogalusa Heart Study, Alzheimer's Disease Neuroimaging Initiative) have implemented various digital voice assessments as part of their cognitive assessment protocols. A Scotland study is dedicated to understanding how speech markers from recording of everyday speech can be used to evaluate brain health [42^▪▪]. Clinical trials are increasingly using voice-based technology, with the appreciation of reducing barriers to participation and increasing representativeness of participant populations [43^▪▪].

OTHER DIGITAL MEASURES FOR ALZHEIMER'S DISEASE DIGITAL BIOMARKER CONSIDERATION

Separate from speech, other platforms, modalities, and measures have been developed and reported as potential digital biomarkers. Qi et al. provide a comprehensive review that gives the illusion that much research has been done in AD digital biomarkers [44]. But similar to digital voice, this review does not check for biomarker validation. Described below are studies that relate digital measures to underlying AD pathological indicators and serve as close approximations to the type of research needed to validate digital phenotypes as digital biomarkers.

Digital metrics obtained from a clock drawing test using a digital pen have been correlated with brain MRI metrics of volume loss [45^▪], and increased PET amyloid and tau [46,47]. Other digital cognitive assessments administered via a computer, tablet, or smartphone have also been associated with AD pathology at the preclinical stage. Digital cognitive metrics, especially memory, have been associated with amyloid deposition measured via PET [48–53]. Moreover, lower performance on digital memory tasks correlated with increasing tau burden [51,54^▪▪]. Individuals with PET AD biomarker profiles have been shown to have worse cognitive profiles than those with no or only one pathology present [55^▪]. Studies with blood-based biomarkers and digital cognition are very limited. One study reported that digital cognitive performance tracked with plasma ptau181 levels [56^▪]. Another study assessed long-term forgetting through remote collection of digital data across seven consecutive days and found lower performance were related to plasma ptau181 [57].

Other digital measures of cognitively related behaviors have also been linked to AD pathology. One study noted area-under-the-curve predictive of PET amyloid from wearable sensors tracking lifestyle behaviors alone (0.70) was nearly comparable to the full-scale model of demographics, questionnaire and wearables (0.79). A virtual reality measure of spatial navigation was associated with elevated amyloid assessed by CSF [58^▪▪]. Other non-cognitive biosensor measures have also been related to AD pathology such as sleep [59^▪,60,61], and heart rate variability measures [62], and include validation with post-mortem AD [63].

OPEN-SOURCE QUALITY MANAGEMENT SYSTEM: THE MISSING LINK TO DIGITAL ALZHEIMER'S DISEASE BIOMARKERS

A quality management system (QMS) is a structured framework of policies, procedures, and resources that organizations use to ensure their products or services consistently meet customer and regulatory requirements while continuously improving their operations. At its core, most QMSs includes document and records management, process management, performance monitoring and measurement, and continuous improvement.

Documentation is a cornerstone of an effective QMS. Table 1 outlines some, but not all, source documentation that need to be considered when conducting research activities that are aimed to meet regulatory requirements. These documents need to be pieced together to adequately cover an entire portfolio of research operations to instill controls of quality and the confidence of partners and regulatory agencies, while simultaneously being lightweight and adaptable enough for any institution or organization to implement.

Table 1.

Illustrative list of QMS quality & procedure regulations, frameworks & considerations

Issuing organization	Document
CFR1	21 CFR Part 11 Electronic Records and Electronic Signatures, 1997
ISO2	ISO-9001:2015: Quality management systems – Requirements
ISO	ISO-19011:2018: Guidelines for auditing management systems
ICH3	ICH E6(R2): Good Clinical Practice: Integrated Addendum to ICH E6(R1)
ICH	ICH E8(R1): General Considerations for Clinical Studies
ICH	ICH E9: Statistical principles for clinical trials – scientific guideline
ISPE4	GAMP5 v5: A Risk-Based Approach to Compliant GxP6 Computerized Systems
NIST7	NIST 800-53: Security and Privacy Controls for Information Systems and Organizations
NIST	NIST 800-171: Protecting Controlled Unclassified Information in Nonfederal Systems and Organizations
FDA8	FDA Guidance on Electronic Source Data in Clinical Investigations
FDA	FDA Guidance on Digital Health Technologies for Remote Data Acquisition in Clinical Investigations
FDA	Framework for FDA's Real World Evidence program

CFR, Code of Federal Regulations; FDA, Food and Drug Administration; GAMP, Good Automated Manufacturing Practice; GxP, Good × Practice; ICH, International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use; ISO, International Organization for Standardization; ISPE, International Society for Pharmaceutical Engineering; NIST, National Institute of Standards and Technology.

There are a variety of loosely defined pathways for FDA approval depending on the use case of the intended Digital Biomarker. Today, meeting FDA approval would consist of parsing through dozens of documents to identify what is relevant for a study before documenting and implementing. Even then, if certain steps are skipped, such as maintaining document control and audit logs, or critical documents missed during this discovery period, there would be no pathway to obtaining regulatory approval. As it currently exists, this is an unrealistic expectation for academics to follow. This is why it is imperative to create an open QMS within the academic environment. This QMS would be a consolidated product consisting of a Quality Manual and Standard Operating Procedures, considering all relevant recommendations and frameworks released by regulatory bodies to provide coverage for all research activities that may contribute to an FDA approved digital biomarker.

If academic research studies can conduct their human studies in QMS environments that are comparable to what pharmaceutical and biotech companies routinely use, it will result in a level of transparency that will solve the current replicability and reproducibility problem, which in turn will then translate more rapidly into clinically useful results.

Currently, the average number of years it takes to bring a new drug to market within the US is about 10–15 years within the pharmaceutical industry [64], which is likely a significant contributing factor to the 17-year research-clinical care gap [65]. But this timeline is much longer for neurological disorders because it has been difficult to elucidate the underlying biological pathways [66^▪▪]. For example, prior to FDA approval of Aducanumab, Lecanemab and Donanemab in 2021, 2023, and 2024, respectively, the last AD drug approved for treatment of those at the moderate to severe stage was Memantine in 2003 and for mild to moderate AD, Aricept in 1996. However, the effectiveness of these drugs, approved 27+ years after Aricept, is relatively modest [66^▪▪]. They slow cognitive decline but do not stop it. This suggests that there are other factors that are yet unknown that must be elucidated before a cure for AD can be realized. The way a QMS system is implemented in academic research could potentially affect the timeline for defining digital biomarkers and translating results into treatments.

IMPACT

If AD human studies could collect and analyze their data in a regulatory-compliant QMS environment, longstanding concerns of replicability and reproducibility may be largely eliminated. Further, direct access to FDA and other regulatory compliant databases will drive down the cost of innovation for the private sector because they can utilize these same data to generate the evidence required for regulatory review. The net result would be an acceleration in scientific discovery through increased competition and a shrinking of the current 17-year research-implementation lag.

CONCLUSION

Receptivity to the idea of AD biomarkers is high. Currently, AD digital phenotypes are mischaracterized as “digital biomarkers.” While digital phenotypes hold scientific and clinical value, they are not biomarker equivalents. A regulatory-compliant QMS provides a potential framework to move the myth of digital biomarkers to reality.

Acknowledgements

The authors thank Dr Shibeshih Belachew for his initial insights on the importance of FDA compliant QMS.

Financial support and sponsorship

This work was informed by research supported by the American Heart Association (20SFRN35360180), the National Institute on Aging (2P30AG013846, AG062109, AG068753, AG072654), the Alzheimer's Drug Discovery Foundation (201902–2017835), Gates Ventures, and the National Science Foundation (EECS-2317079, CCF-2200052, DEB-2433726).

Conflicts of Interest

There are no conflicts of interest. Unrelated to the work reported, RA is a scientific advisor to Signant Health and NovoNordisk.

REFERENCES AND RECOMMENDED READING

Papers of particular interest, published within the annual period of review, have been highlighted as:

• ▪ of special interest

• ▪▪ of outstanding interest