What is wearable data science?
Wearable data science is the engineering discipline of extracting, processing, and validating meaningful signals from wearable sensor data for clinical and commercial health applications. Where consumer wearable analytics tolerates approximation, clinical wearable data science requires that every pipeline component be deterministic, reproducible, and regulatorily defensible. Raw accelerometer, gyroscope, and photoplethysmography signals from devices like Apple Watch, Empatica, Samsung, and ActiGraph pass through preprocessing pipelines that normalise sampling rates, axes, and units before clinical algorithms compute their outputs. Cross-platform validation — confirming that the same algorithm produces equivalent results across multiple devices using different sensor hardware — is a core technical challenge distinct from standard data science work. Engagements typically cover signal processing pipeline development, algorithm portability across hardware platforms, and output validation against clinical reference standards. Devsort has built and validated cross-platform wearable pipelines for PKG Health (now Empatica) covering five device platforms under FDA, CE, and TGA requirements.
How does clinical wearable data science differ from consumer analytics?
Consumer wearable analytics — step counts, sleep summaries, resting heart rate trends — is built on aggregated population statistics where individual errors average out. Clinical wearable data science operates on individual patient data where every sample matters. A misclassified movement event, an incorrectly normalised accelerometer reading, or a sampling artefact that goes undetected can produce a clinically incorrect output for a specific patient at a specific time.
The second distinction is reproducibility. A clinical algorithm must produce the same output for the same input every time, across different hardware revisions, different software environments, and different processing platforms. Building a preprocessing pipeline that guarantees this across five device types — each with different sensor hardware, different sampling rates, and different axis conventions — is the core technical challenge of clinical wearable data science.
Devsort's experience comes from building exactly this kind of pipeline for PKG Health's movement disorder monitoring system. The work required understanding the signal characteristics of each device at the hardware level before a single line of algorithm code could run on it.
Cross-platform work: the PKG Health engagement
PKG Health's movement disorder algorithms were originally validated on a single hardware platform. Expanding the algorithm suite to run on Apple Watch, Samsung Galaxy Watch, Sony devices, Empatica E4, and ActiGraph required building a preprocessing pipeline from scratch — normalising the raw output of five fundamentally different accelerometers to a common input format that the clinical algorithms could trust.
Each device has different axis orientation conventions, different sampling frequencies, different units, and different filter characteristics baked into the hardware. The preprocessing pipeline detects which device produced the data, applies the appropriate transformation, resamples to the algorithm's expected input frequency, and outputs a standardised signal. Every transformation was validated against clinical datasets to confirm that algorithm outputs were equivalent across platforms.
The same engagement also involved reimplementing a subset of the validated Python algorithms in embedded C for deployment on the Empatica wrist device — a further cross-domain step from cloud processing to embedded real-time execution.
Wearable platforms we have worked with
Frequently asked questions
How does signal processing differ between consumer and clinical wearables?
Consumer signal processing optimises for battery life and user experience — data is aggregated, smoothed, and displayed as trends. Clinical signal processing must be deterministic and lossless at the sample level: every raw accelerometer or PPG reading must be preserved, validated, and processed in a documented, reproducible way. The downstream algorithm's output is only as trustworthy as the preprocessing that precedes it.
Which wearable platforms can your algorithms run on?
We have validated preprocessing pipelines and algorithm ports for Apple Watch, Empatica E4, Samsung Galaxy Watch, Sony SmartWatch, and ActiGraph devices. For new platforms, we assess the hardware's sensor specifications, build the preprocessing transformation, and validate output equivalence against a reference device before certifying that the algorithm runs correctly on it.
What does cross-platform validation require?
Cross-platform validation requires a reference dataset collected simultaneously on the target device and a reference device, a preprocessing pipeline that normalises both to a common format, and a statistical comparison of algorithm outputs across both. Any differences are investigated at the signal level to determine whether they are within the tolerance defined by the regulatory or clinical specification. The validation process is documented in full.
How long does a cross-platform wearable data project take?
A single-platform extension — adding one new device to an existing validated algorithm — typically takes six to twelve weeks: two to four weeks for sensor characterisation and preprocessing, two to four weeks for algorithm porting and initial validation, and two to four weeks for full validation and documentation. Multi-platform programmes covering three or more devices run three to nine months depending on the number of algorithms and the complexity of each device's signal characteristics.
Can you work with custom or proprietary wearable hardware?
Yes, provided we have access to the hardware specifications and raw data output format. Custom hardware typically requires more upfront sensor characterisation work than a commercial device with published specifications, but the pipeline-building and validation process is the same. If the hardware is a prototype, we work with engineering teams to establish the calibration and output format conventions before building the preprocessing layer.
Do you work on non-clinical wearable applications?
Yes. Our core expertise is clinical-grade pipelines, but the signal processing and cross-platform techniques apply equally to fitness, occupational health, and research wearable applications. The difference is validation rigour: clinical applications require documented equivalence against regulatory standards; research applications require documented methodology; fitness applications require performance benchmarking. We scope each engagement against the appropriate standard for the application.