The Digital Health Technology Field GuideSensor data in research, regulation, and care, with a close look at accelerometer and voice data. Boyce Data Science, snapshot dated September 2026.

How a sensor reading becomes evidence

The same chain of steps sits behind a step count on a watch, a speech score in an amyotrophic lateral sclerosis (ALS) study, and a primary endpoint in a Duchenne muscular dystrophy trial, and many disagreements in this field turn out to be about one specific link in it.

1. Sensor and raw signal. A device samples acceleration, sound pressure, light absorption, or another physical quantity many times per second and writes it to a file.
2. Algorithm. Software turns the raw signal into something interpretable, such as steps, sleep periods, fundamental frequency of the voice, or pause length.
3. Digital measure. The algorithm output is summarized over a defined window, for example average daily steps across valid wear days, or speaking rate in a standard reading passage.
4. Endpoint, biomarker, or outcome assessment. The measure is given a role in a study or in care, with a stated concept of interest and context of use. Regulators use the terms digital biomarker and electronic clinical outcome assessment (eCOA) for different roles, and the glossary explains the distinction.
5. Decision. A person, a trial, a regulator, or a payer acts on the result, and the strength of evidence needed depends on that decision.

How to use this guide

The guide shows one stage at a time. The bar under the title jumps to any stage, and the buttons at the foot of each stage move forward and back. The search box looks across every stage, including the collapsed lookup tables.

Entries for data sets, software, organizations, guidance documents, and funders each have an Add to shortlist button. Those entries go into your shortlist, where you can add notes and download the list as Markdown, as a comma-separated values (CSV) file, or as a save file that reloads your work in a later session. Nothing you enter leaves your browser.

This is a snapshot dated September 2026. It has not been peer reviewed and it is not regulatory, legal, or medical advice. The field changes quickly, so entries link to the primary source, and the literature and funding stages include live searches you can rerun to see what has changed since the snapshot.

Movement

Wrist actigraphy began in sleep research in the 1970s, and scoring rules published in the early 1990s are still in use. Exercise science adopted waist-worn count-based monitors in the 1990s, and national surveys followed. Storage and battery gains in the 2010s let devices keep the raw signal, and that change made open algorithms and very large cohorts possible. UK Biobank and the National Health and Nutrition Examination Survey (NHANES) are the reference points many analysts learn on.

Voice and speech

Acoustic analysis of speech dates to the sound spectrograph of the 1940s, and speech pathology built a detailed perceptual vocabulary for motor speech disorders well before computers could measure the same features. Free software such as Praat spread acoustic measurement widely. Work on Parkinson's disease in the late 2000s showed that recordings made at home could track symptoms, and the machine learning community added standard feature sets, shared challenges, and later self-supervised speech models.

Regulation and shared frameworks

Regulators first met these technologies as medical devices and only later as measurement tools inside drug trials. The European Medicines Agency (EMA) qualified a wearable-derived endpoint in 2019 and again as a primary endpoint in 2023. FDA published final guidance on DHTs for remote data acquisition in December 2023 and has continued with workshops, a request for information, and a funding opportunity in 2026.

What an accelerometer records

A triaxial accelerometer reports acceleration along three perpendicular axes in units of gravity (g). A device at rest still reads about 1 g in total because gravity is always present, which is why orientation can be estimated from a still device and why most processing begins by separating gravity from movement.

• Sampling rate. Research devices commonly record between 30 and 100 samples per second (hertz, Hz). UK Biobank used 100 Hz and NHANES 2011 to 2014 used 80 Hz. Consumer devices often sample at similar rates internally and share only minute-level or daily summaries.
• Dynamic range. The range is commonly plus or minus 8 g. A narrow range clips vigorous movement.
• Wear location. Wrist, hip, thigh, lower back, and ankle each suit different measures. Wrist wear tends to have better adherence in published studies, thigh wear separates sitting from standing well, and lower back or ankle wear is preferred for gait, so a measure validated at one location should not be assumed to hold at another.
• Raw data and epoch summaries. Older and consumer systems store a summary per epoch, often a minute. Raw data let you rerun the analysis when algorithms improve, at the cost of files that reach hundreds of megabytes per person per week.
• Companion sensors. Gyroscopes, magnetometers, temperature, and light sensors often sit in the same device. A unit with accelerometer and gyroscope is usually called an inertial measurement unit (IMU).

What a voice recording holds

A digital recording is sound pressure sampled many thousands of times per second. Quality is set at the moment of recording by the microphone, its distance from the mouth, the room, and any compression applied by the app or phone, and none of these can be repaired afterward.

• Sampling rate and bit depth. Acoustic voice analysis commonly uses 44.1 or 48 kilohertz (kHz) at 16 bits. Many speech recognition models resample to 16 kHz, and telephone audio is limited to about 8 kHz, which removes high-frequency detail.
• Compression. Uncompressed or lossless files preserve the waveform. Lossy formats discard information, and published comparisons report that perturbation measures such as jitter and shimmer are among the features most affected, so check the literature before pooling recordings made with different codecs.
• Tasks. Common tasks include a sustained vowel, rapid syllable repetition (the diadochokinetic or DDK task, often "pa-ta-ka"), a standard reading passage, picture description, and free speech. Each task stresses a different part of the system, so a protocol should state which construct each task is meant to reach.
• Identifiability. A voice recording can identify a person and can include other people's speech. Many programs therefore share derived features openly and keep raw audio behind a controlled-access agreement, as Bridge2AI-Voice does.

The movement pipeline

1. Read and check. Confirm sampling rate, duration, clock, and gaps, and plot a few days before computing anything.
2. Calibrate. Auto-calibration uses still periods, when the vector magnitude should equal 1 g, to correct small sensor errors.
3. Detect non-wear. Rules look for long periods with almost no variation. How non-wear is handled (dropped, imputed, or modeled) should be decided before unblinding, because non-wear is often related to health.
4. Summarize acceleration. Common summaries include Euclidean norm minus one (ENMO), mean amplitude deviation (MAD), activity counts, and MIMS units.
5. Classify. Algorithms label sleep, sedentary time, walking bouts, steps, or postures, using thresholds or machine learning models.
6. Aggregate. Person-level measures are built over valid days, for example average acceleration, time above an intensity threshold, daily steps, cadence, real-world walking speed, or rest-activity rhythm measures.

The voice and speech pipeline

1. Check the recording. Listen to a sample, check clipping, background noise, and whether the task was done as prompted.
2. Segment. Voice activity detection and, where needed, speaker diarization separate the participant from silence and from other speakers.
3. Extract acoustic features. These include fundamental frequency, jitter, shimmer, harmonics-to-noise ratio, cepstral peak prominence, formants, and MFCCs.
4. Extract timing features. Speaking rate, articulation rate, pause frequency and duration, and syllable rate and regularity in DDK tasks.
5. Transcribe and extract language features. Automatic speech recognition gives transcripts for lexical, syntactic, and semantic measures. Recognition error itself has been studied as a proxy for intelligibility.
6. Model. Interpretable features, learned embeddings, or both feed statistical or machine learning models, with evaluation split by person and not by recording.

United States: FDA

The December 2023 final guidance, Digital Health Technologies for Remote Data Acquisition in Clinical Investigations, is the central document for trials. It covers choosing a DHT that is fit for purpose, verification and validation, endpoints, risks, data retention, and sponsor and investigator roles. It sits within commitments made under the seventh Prescription Drug User Fee Act reauthorization (PDUFA VII), which also produced a March 2023 framework, a cross-center DHT Steering Committee, public meetings, and demonstration projects.

In 2026 FDA issued a request for information on advancing DHTs in drug and biologic trials, with comments due June 1, 2026, opened a funding opportunity from July 20 to August 20, 2026, and held a public workshop on statistical considerations for digitally derived endpoints on August 27, 2026. Separately, on January 6, 2026, FDA revised its general wellness and clinical decision support guidances, and legal commentators describe the general wellness revision as widening enforcement discretion for non-invasive wearables that report physiologic values for wellness uses.

Routes for a measure to gain standing beyond a single trial include the Drug Development Tool qualification programs for COAs and biomarkers, the Innovative Science and Technology Approaches for New Drugs (ISTAND) program, and the Medical Device Development Tools (MDDT) program. A 2026 systematic review reported that no real-world walking speed measure had been qualified by FDA at the time of writing, and it found letters of intent on FDA's website.

Europe and the United Kingdom

EMA's qualification of novel methodologies procedure produced the field's reference case. SV95C, measured at the ankle by a suitable wearable device, was qualified as a secondary endpoint for Duchenne muscular dystrophy in 2019 and in July 2023 as a primary endpoint in superiority studies for ambulant patients aged four and above, as an alternative to the six-minute walk test, provided it is supported by consistent findings in established secondary endpoints. EMA issued a letter of support to the Mobilise-D consortium in April 2020 for digital mobility outcomes as monitoring biomarkers in Parkinson's disease.

Devices and software with a medical purpose fall under the European Union Medical Device Regulation, and the Artificial Intelligence Act and the European Health Data Space regulation add obligations for high-risk artificial intelligence (AI) systems and for secondary use of health data. The General Data Protection Regulation treats health data, and biometric data used to identify a person, as special categories. In the United Kingdom, the National Institute for Health and Care Excellence (NICE) publishes an evidence standards framework for digital health technologies, and Germany's digital health application (DiGA) pathway offers reimbursement for listed apps.

Payment and privacy in the United States

The 2026 Medicare Physician Fee Schedule, finalized in November 2025 and effective January 1, 2026, added remote monitoring codes for shorter periods, covering 2 to 15 days of data in a 30-day period and 10 minutes of management time, alongside the existing codes that require 16 days and 20 minutes. In February 2026 the Centers for Medicare and Medicaid Services (CMS) announced a Medicare App Library within its health technology ecosystem work.

The Health Insurance Portability and Accountability Act (HIPAA) applies to covered entities and their business associates, so data held by a consumer wearable company are often outside it. The Federal Trade Commission's Health Breach Notification Rule, state laws such as Washington's My Health My Data Act, and biometric privacy laws such as the Illinois Biometric Information Privacy Act, which names voiceprints, cover parts of that gap.

In clinical care

• Remote monitoring programs. Medicare and many private payers reimburse RPM and RTM, and the 2026 fee schedule added codes for shorter monitoring periods. Programs usually need a defined population, a response protocol for out-of-range values, and staff time to review data.
• Patient-generated data in the record. Major EHR systems can receive consumer device summaries through patient portals and FHIR interfaces. Many clinicians report that unfiltered streams add workload, so programs often agree in advance which summaries are reviewed and when.
• Consumer features with device authorization. Some smartwatch features, including irregular rhythm notifications, ECG recording, and sleep apnea notifications, have FDA marketing authorization, and general wellness features do not. The labeling states the intended population and what the feature does not do.
• Seizure detection. Wrist-worn devices that detect convulsive seizures and alert caregivers have FDA clearance, and an international clinical practice guideline published in 2021 addresses their use for generalized tonic-clonic seizures.
• Speech and communication. For people with ALS and other conditions that affect speech, voice banking and message banking preserve a personal voice for later use in a communication device, and projects such as the Speech Accessibility Project and Project Euphonia aim to improve recognition of atypical speech.

Patient-centered use

• People living with a condition often want their own data returned in a form they can read, and want to know who else can see it. Plans for returning individual data and study results are best made with participants during design.
• Burden belongs in the design, since charging, syncing, skin irritation, and repeated tasks all reduce adherence, and the people with the most advanced disease are often the first to drop out, which biases results.
• Access is uneven, because device cost, broadband, smartphone model, language, dialect, and disability all affect who can take part and how well algorithms perform for them.
• Monitoring can feel like surveillance, so continuous sensing in the home, and any microphone, should come with clear limits on what is recorded, who reviews it, and how to pause it.

Regulatory methods are becoming specific

FDA's 2026 activity moved from general principles toward methods: statistics for digitally derived endpoints, a technical specification for one continuous sensor stream, and funding aimed at comparing digital with traditional measures. The current user fee agreement runs through fiscal year 2027, so commitments for the next cycle are likely to be discussed publicly during the coming year.

Composite and multimodal measures

DiMe lists a current project on combining digital measures into composite and multimodal endpoints, and studies such as the ALS Therapy Development Institute's work with Google combined voice and accelerometer models within one cohort. The statistical and regulatory treatment of composites built from different sensors is still unsettled.

Large pretrained models for sensor data

Self-supervised models trained on very large wearable cohorts, and pretrained speech models, are replacing hand-built features in some tasks. They reduce the labeled data needed and raise new questions about interpretability, about performance for people who are unlike the training population, and about how to lock a version for a trial.

Ethically sourced voice data at scale

Bridge2AI-Voice released successive adult versions through 2025 and 2026, added a pediatric release, and opened raw audio through a controlled process. Its split between openly available derived features and controlled raw audio may become a common model. Voice cloning technology also increases the consequences of any leak of raw audio, which strengthens the case for that split.

Contactless and ambient sensing

FDA's 2026 funding notice gives contactless room sensors for apnea in pediatric patients as an example. Sensing that needs nothing worn suits infants, people with advanced disease, and people with cognitive impairment, and it brings consent questions for everyone else in the room.

Wider wellness space and changing payment

The January 2026 general wellness revision and the 2026 Medicare remote monitoring codes both widen the room for sensor-based products in everyday care. Whether evidence generation keeps pace is an open question that independent evaluators and patient groups are watching.

Rare disease and pediatric measures

After SV95C, DiMe's collaborative community launched a core set of digital clinical measures for pediatric rare diseases. Small populations make shared measures and shared data especially valuable, since no single sponsor can validate a measure alone.