Revitalyze

Landing page

Inspiration

We wanted to explore whether a device as common as a smartphone could surface meaningful health insights without requiring specialized hardware. Many important vital signs are still difficult to access continuously, especially for people without easy access to medical equipment or routine screening. That made us interested in building a system that could turn an everyday camera into a lightweight physiological sensing tool.

We were especially drawn to photoplethysmography (PPG), since it offers a way to recover pulse-related information from subtle color changes in fingertip video. From there, we asked a harder question: could we go beyond simply estimating heart rate and instead build a richer pipeline that combines signal processing, machine learning, and LLM-based interpretation? That question became Revitalyze.

What it does

Revitalyze is a multimodal health-insight pipeline built from three main layers:

Signal extraction from smartphone video.
We capture a short fingertip video and recover a PPG waveform from the central region of interest. From this waveform, we estimate heart rate and a signal-confidence score that reflects how clean or noisy the recovered pulse signal is.
Prediction of additional physiological metrics.
On top of the recovered biosignal, we use a custom-trained PyTorch model to predict higher-level health indicators such as:
- respiration rate,
- blood oxygen,
- blood pressure,
- and additional metrics described in the README.
LLM-powered interpretation.
We pass these structured outputs into Lava, which lets us query an LLM to generate concise, non-diagnostic guidance for the user. Instead of pretending to provide a medical diagnosis, the system explains the results, highlights possible concerns, and suggests when a user may want to follow up with a clinician.

In short, our pipeline moves from raw pixels ➔ biosignal ➔ physiological estimates ➔ human-readable guidance.

How we built it

Our backend is built in Python with Flask and OpenCV. We process the uploaded fingertip video by:

detecting a central crop of the frame,
averaging color values over time,
filtering candidate color traces,
selecting the best pulse waveform using a quality heuristic,
estimating heart rate from the recovered signal,
and assigning a confidence label based on agreement between spectral and peak-based estimates.

We then combine those outputs with our learned physiological predictors. Our ML component was trained on BUTPPG and PULSEDB, and we also consulted prior literature and public physiological datasets during development, ensuring our approach is grounded in established physiological modeling and supported by prior validated research, as referenced in our README.

Finally, we integrated Lava as the gateway for our LLM layer. That lets us feed the model structured health metrics and receive concise, user-facing recommendations while keeping the system modular and easy to extend.

Intellectual depth

A big part of this project’s complexity is that it spans multiple technical layers that each fail in different ways:

Computer vision / signal extraction: recovering a clean pulse waveform from noisy phone video is hard because motion, pressure changes, lighting shifts, and camera auto-exposure can all corrupt the signal.
Time-series processing: we had to design filtering, smoothing, and peak-selection logic that was robust enough to recover meaningful physiological structure instead of just visual noise.
Machine learning: mapping biosignals to downstream physiological estimates required training and validating predictive models rather than relying on a single heuristic.
Interpretability: raw predictions are not useful unless users can understand them, so we added an LLM layer to translate technical outputs into understandable, non-diagnostic feedback.
Trust and uncertainty: instead of hiding uncertainty, we explicitly expose signal confidence so users can tell when the measurement quality is low.

That combination of sensing, inference, and explanation is what makes Revitalyze more than a simple heart-rate app.

Challenges we faced

One of our biggest challenges was signal quality. Recovering a stable PPG waveform from fingertip video is extremely sensitive to motion, inconsistent finger placement, and exposure changes. Even when a waveform looked promising, heart-rate estimates could drift if the signal was noisy.

Another challenge was responsibility. Because this is a health-related project, we were careful not to frame the output as diagnosis. We designed the final recommendation layer to be cautious and supportive rather than definitive, so the system acts more like an intelligent screening or wellness assistant than a medical authority.

We also had to bridge very different technical components built by different teammates: video-based PPG extraction, learned predictors for additional metrics, and LLM-based summarization. Making those pieces work together coherently in a single user-facing flow was one of the hardest and most rewarding parts of the build.

What we learned

We learned that building health technology is not just about getting a prediction — it is about handling noise, uncertainty, interpretability, and user trust all at once. We also learned how powerful it can be to combine classical signal processing with machine learning and LLMs, especially when each layer does a distinct job instead of trying to solve everything alone.

Most importantly, we learned that accessibility matters. If health sensing can be done with only a phone camera and an intelligent backend, then early insight and screening tools can become far more widely available.

Societal impact

Revitalyze lowers the barrier to basic health insight by using a commodity device that millions of people already have: a smartphone. By combining noninvasive sensing, learned physiological prediction, and explainable AI, our system points toward a future where early health screening is more accessible, scalable, and interpretable.

While Revitalyze is not a diagnostic tool, it shows how low-cost sensing and AI can help people notice trends earlier, understand their bodies better, and decide when to seek professional care.