From physiological signals to product-ready monitoring
Leading applied R&D across ECG, PPG, temperature and oxygen-saturation data, from collection and evaluation to Android and cloud integration.
Case record
The same seven questions, every time.
- Context
- Point-of-care monitoring across ECG, PPG, temperature and oxygen-saturation streams that differ in sampling, artefacts and interpretation.
- My role
- Data Scientist / R&D Project Lead — led the applied R&D workflow from signal preparation through model development and evaluation.
- Constraint
- A single modelling step could not carry the whole system, and outputs still had to move into Android and cloud environments.
- Decision
- Treat integration as part of the research problem, and build reproducible experiments that preserve the conditions under which each result was produced.
- Artefact
- Repeatable collection and signal-processing workflows, evaluated time-series classification approaches, documented assumptions, and a supported path to Android and cloud integration.
- Result
- 94% ECG classification accuracy in project evaluation, and a disciplined route from raw signal to evaluated component that kept performance, integration and limits visible together.
- Boundary
- The accuracy figure belongs to the project protocol and dataset. It is not a diagnostic, clinical-performance or deployment-wide claim.
Evidence ledger
What the work can show.
- Signals
- ECG · PPG · temperature · SpO₂Different sampling and artefact profiles were handled as an end-to-end workflow.
- Evaluation
- 94% project accuracyA result within its project protocol—not a diagnostic or clinical-performance claim.
- Delivery path
- Android and cloud integrationIntegration constraints were treated as part of the applied research problem.
The challenge
Point-of-care monitoring has to turn noisy physiological signals into information that can support a usable product. ECG, PPG, temperature and oxygen-saturation streams differ in sampling, artefacts and interpretation, so a single modelling step cannot carry the whole system.
The work required an end-to-end view: how data was collected, how signals were filtered, what the models were evaluated against, and how outputs would move into mobile and cloud environments.
Developing the signal workflow
I led the applied R&D workflow from signal preparation through model development and evaluation. Reproducible experiments made it possible to compare decisions without losing the conditions under which a result had been produced.
Integration was treated as part of the research problem. Supporting Android and cloud delivery exposed practical constraints early and helped keep the analytical work connected to the intended monitoring experience.
- Defined repeatable collection and signal-processing workflows.
- Developed and evaluated time-series classification approaches.
- Documented assumptions and experiment conditions.
- Supported the path from model outputs to Android and cloud integration.
Evidence and limits
Project evaluation reached 94% ECG classification accuracy. That figure is useful evidence within its evaluation setting, not a substitute for describing the data, protocol and intended operating context.
The more durable outcome was a disciplined route from raw signal to evaluated component—one that made performance, integration and limitations visible together.
Connected evidence
Where this sits in the rest of the work.
- Point-of-care intelligence in Selected SystemsThe homepage system entry, with its evidence instrument and boundary.
- Published workThe first-author ECG sample-length study that sits alongside this programme.
- Data quality and model reliability labA working demonstration of why signal quality has to be part of evaluation, not cleanup afterwards.