Bryan Bulte, President @ Sensor Bio | Health Intelligence Platform | Wearables & Human Performance, recently shared something that made me stop scrolling: "PPG is quietly becoming one of the most valuable signals in healthcare. But most wearables were never built to capture it properly." He followed that with a detail that should matter to anyone building in digital health: researchers have now released "one of the largest curated PPG datasets ever assembled," with millions of 30-second segments tied to real clinical outcomes.

That combination - clinical labels at scale and a sensor signal that many people still treat as a fitness-only metric - is exactly why Bulte's post resonated. If PPG is going to graduate from step-count-adjacent insights to serious clinical and research use cases, the industry has to get honest about a simple constraint: you cannot learn what you did not capture.

PPG is not a "fitness signal"

Bulte's core claim is blunt: "PPG isn’t a 'fitness signal.' It’s a physiological signal." That distinction matters because it changes what you optimize for.

A fitness signal is often good enough when it is directionally correct. It can be smoothed, summarized, and turned into a score that feels consistent day to day. A physiological signal is different. If you want to infer autonomic balance, cardiovascular dynamics, respiration, recovery, or subtle deviations from baseline, you need the underlying waveform quality to be high and the data to be preserved.

"Once the signal is smoothed, compressed, or scored too early, those possibilities disappear."

PPG (photoplethysmography) measures changes in blood volume in the microvascular bed of tissue, typically using an optical sensor. That waveform contains information about timing, amplitude, and shape. Those details are where nuance lives, like beat-to-beat variability proxies, respiratory modulation, peripheral vasoconstriction patterns, and responses to stressors.

Why the new large-scale dataset is a big deal

When Bulte highlights "millions of 30-second PPG segments tied to real clinical outcomes," he is pointing to a shift from toy datasets and small validation studies to something that can support robust modeling.

Large curated datasets do three important things:

They make it possible to benchmark algorithms across diverse populations and conditions rather than a narrow lab cohort.
They reduce the temptation to overfit, because scale exposes brittle shortcuts.
They invite research beyond heart rate, into multi-factor physiology.

It also raises the bar for wearable hardware and data access. If the research community can train and validate models on high-resolution segments, the next question becomes operational: can a real-world device collect comparable quality data consistently, across motion, skin tones, temperatures, and daily life?

What high-quality PPG can reveal (and why it is expanding)

Bulte lists a set of domains researchers are already exploring with clean, high-resolution PPG:

Hydration status and fluid shifts
Inflammation and physiological stress load
Sympathetic / parasympathetic dominance
Sleep architecture and recovery depth
PTSD, anxiety, and chronic stress patterns
Illness onset and early deviation from baseline

If you have worked with PPG in practice, you know the pattern: every time signal quality improves, the "feature space" expands. You move from simple averages to variability, from daily summaries to within-night sleep staging signals, from a single metric to multi-dimensional trajectories.

Two ideas are worth emphasizing here.

1) Baselines beat population averages

Many of the most valuable applications Bulte hints at are not about comparing you to the average person. They are about comparing you to you, over time. Early illness detection, chronic stress patterns, and recovery depth all benefit from longitudinal history. That is why capturing consistent raw or minimally processed data matters: you want the same underlying measurement so changes are interpretable.

2) Context turns a signal into an insight

Bulte also notes that when high-quality PPG is paired with "motion, temperature, respiration, ECG, sleep timing, demographics, and longitudinal history" the insight space grows "exponentially."

That rings true because many confounders are contextual. Motion artifacts can mimic physiological changes. Temperature shifts can change peripheral perfusion. Sleep timing affects autonomic state. If you have the contextual streams, you can separate artifact from physiology and build models that generalize.

The uncomfortable truth: most wearables were not designed for this

Bulte answers his own question: "So why don’t most consumer wearables give researchers this level of data?" His take is that it is not negligence. It is tradeoffs.

He lists practical constraints that every product team recognizes:

Higher sampling rates drain battery
Devices with interactive displays must allocate battery to the screen
Continuous raw capture increases storage and compute costs
Smoothing reduces noise and user complaints
Early compression into "scores" simplifies UX

These are reasonable decisions if your primary goal is a consumer-friendly device that feels stable, lasts days, and presents clean dashboards. But they collide with research-grade needs.

Here is the key sentence Bulte repeats in different forms: once you smooth, compress, or score too early, the information loss is permanent. In other words, you can always summarize later, but you cannot "unsummarize" after details have been removed.

Signal fidelity is a product philosophy, not a feature

What I found most useful in Bulte's post is that he frames PPG quality as an architectural choice made early. He says Sensor Bio chose "9 years ago" to focus relentlessly on PPG fidelity, with decisions like:

Designed for signal fidelity over battery marketing
Engineered to capture physiology, not just trends
Sampling and filtering choices driven by data integrity, not UI simplicity
Architecture built to support future algorithms, not just current features
Data structured for longitudinal insight, not daily scores

That list reads like a philosophy: build for future discovery, not just current widgets.

This is the strategic bet: algorithms will mature, clinical use cases will expand, and datasets like the one he references will accelerate that curve. If you collected only scores, you limited yourself to what you believed at launch. If you collected high-quality waveforms, you left the door open for new models, new endpoints, and new validations.

A practical checklist for teams working with PPG

If Bulte's post sparked a discussion in your organization, here are a few grounded questions to ask. They are not about chasing perfection, but about making the tradeoffs explicit.

Data capture and preservation

What sampling rate do you use for PPG in real-world mode, not just special tests?
Do you preserve raw waveforms, or only filtered signals and derived metrics?
What happens to the data during motion, low perfusion, or poor fit?

Processing and compression

Where do filtering and smoothing occur (on-device, in firmware, on the phone, in the cloud)?
Are you compressing in a way that destroys morphology (shape), timing, or amplitude details?
Can researchers access the pre-score data needed to validate new hypotheses?

Context and longitudinal structure

Are motion, temperature, and sleep timing aligned with PPG timestamps?
Is the data stored in a way that supports longitudinal analysis across months?
Do you track device-side changes (firmware, calibration, sensor revisions) so models do not drift silently?

The new question to ask

Bulte ends with a reframing I think will become more common as clinical datasets grow: "The question is no longer 'Can PPG do this?' It’s 'Can your wearable actually capture the signal required to do it?'"

That is the right question because it pushes beyond marketing claims and into measurement validity. PPG can be extraordinary, but only if the device captures it with enough fidelity and consistency to support clinical-grade inference.

If you want to go deeper, Bulte linked the paper here: https://lnkd.in/gMRRAdHQ

This blog post expands on a viral LinkedIn post by Bryan Bulte, President @ Sensor Bio | Health Intelligence Platform | Wearables & Human Performance. View the original LinkedIn post →