Helios Infrastructure
The Infrastructure
Most health AI is trained on hospital data from wealthy countries and validated on populations that don't look like the people we serve. We built ours differently.
What We Are
Not a product company.
An infrastructure company.
The Health Plus Card, Helios Pulse, and the drone network are entry points into a health intelligence layer — not the layer itself. The layer is what happens when continuous biometric data, clinical records, payment history, and care access patterns for millions of people flow through one system designed for African infrastructure conditions from day one.
Core ML Architecture
Baseline, not threshold.
Western clinical thresholds were derived from studies conducted largely on European and North American populations. A resting heart rate of 60–100 bpm is the textbook normal. It does not account for population-level variation in baseline activity, high-altitude adaptation, genetic variance in autonomic tone, or the dietary patterns common across sub-Saharan Africa.
When a system trained on those thresholds is deployed on a different population, you get two failure modes: false positives that erode user trust and false negatives that miss real clinical events.
Helios Pulse does not compare your readings to a textbook. During the first weeks of continuous wear, the device builds a personalized physiological baseline for each user — resting HR, SpO₂ distribution, HRV range, temperature variance, and activity-adjusted norms. Alerts fire when your readings deviate from your normal, not from a population average.
That is personalized baseline drift detection. The implementation requires enough on-device memory to hold rolling statistical windows per sensor channel, and a lightweight anomaly model that runs inference without a network call. The threshold isn't fixed — it updates continuously as the model learns your physiology.
Signal Pipeline
From skin to alert.
Raw sensor output is not a diagnosis. Here is the actual pipeline between a sensor touching skin and an alert reaching a health worker.
Acquisition
Pulse captures three raw signal types continuously: a photoplethysmography (PPG) waveform via the optical SpO₂/heart-rate sensor, a single-lead ECG waveform via the ECG module, and skin temperature. All three channels run in parallel, not sequentially.
Artifact rejection
Motion artifacts are the primary source of noise in wrist-worn PPG. A person walking, gesturing, or riding a boda-boda introduces signal noise that looks like a heart rate spike if left unfiltered. The accelerometer channel detects motion events and suppresses or flags PPG/ECG readings taken during high motion, rather than reporting a noisy reading as a clean vital sign.
Feature extraction
From the cleaned PPG waveform: heart rate, SpO₂ (via red/infrared absorption ratio), and heart rate variability (HRV) derived from beat-to-beat interval timing. From ECG: R-peak detection, RR-interval sequence, and rhythm classification features. This is where "heart rate" stops being a single number and becomes a time series with variability structure — the input the baseline model actually uses.
Baseline modeling
During an initial observation window, the system builds a per-individual reference distribution for each vital — not a fixed clinical range. Once established, incoming readings are scored against that individual's distribution using deviation magnitude and persistence. A single noisy reading is treated differently from a sustained multi-hour drift.
On-device inference
Scoring runs locally on the wearable's own processor — not in the cloud — so an anomaly can be detected and an alert triggered with zero network dependency. An alert that only fires when a cloud call succeeds is an alert that misses the exact moments connectivity is worst.
Alert dispatch and sync
When a score crosses threshold, the wearer receives an immediate on-device notification, and an alert payload is queued for the registered health worker or doctor. Both dispatch as soon as any network path is available — Bluetooth to a paired phone, then cellular/WiFi from the phone. The full underlying waveform data syncs to the cloud opportunistically afterward for clinical review.
Clinical Validation
The sensors.
Numbers without methodology are marketing. These figures come from validation against clinical gold-standard reference devices. Full validation methodology and study design available on request at centraloffice@helioshealths.com.
ECG accuracy vs. 12-lead clinical ECG
98.5%
SpO₂ accuracy vs. clinical pulse oximetry
96.2%
Heart rate accuracy
99.1%
Alert sensitivity (pre-symptomatic)
94%
False positive rate
3.2%
Architecture
Edge-first by necessity, not by design choice.
Most wearables assume constant connectivity because their markets have it. Ours doesn't. In the environments where Helios Pulse operates — rural clinics, community health worker visits, high-density peri-urban areas with variable cell coverage — assuming a live data connection is an engineering error, not a product decision.
Anomaly scoring runs on-device before any network call. Inference happens on the wearable itself. Alert payloads and scored biometric data sync automatically and opportunistically as soon as any network path appears — Bluetooth to a paired phone, then cellular or WiFi onward. Sometimes that happens within minutes. Sometimes it takes hours. The system doesn't wait. The cloud layer aggregates what the edge has already computed.
This also has a privacy implication: raw waveform data does not leave the device until a network path is available to sync it. What transits the network first is the scored, structured output — the alert payload. The underlying waveform follows opportunistically for clinical review.
Conversational Layer
Ask Helios.
What it is
Ask Helios is a conversational intelligence layer over the platform. It handles product guidance, user onboarding, benefit explanations, and triage-adjacent health Q&A routed to the appropriate care pathway. It understands the Helios product surface — Health Plus Card balances, Pulse data summaries, facility locations, and drug availability through the drone network.
It is built on a retrieval-augmented generation architecture with a tool layer for structured data queries. Phase 2 moves to purpose-trained models with direct access to longitudinal clinical records, enabling deeper health record navigation and clinical decision support.
What it is not
Ask Helios does not touch the clinical inference pipeline that runs on Pulse. The anomaly detection model, the personalized baseline engine, and the pre-symptomatic alert system are separate systems. They do not route through the conversational layer and are not influenced by it.
That boundary is a design decision, not a gap. Clinical inference requires deterministic behavior, audit trails, and a different reliability contract than a conversational interface. Systems that know their own scope are more trustworthy than systems that try to do everything.
Logistics Network
The network as a latency problem.
Rural healthcare facilities don't have a care quality problem or a clinician motivation problem. They have a supply latency problem. Essential medicines, diagnostics, and blood products arrive 3 to 5 days after order, if the road is passable. The drone network is an answer to that specific constraint.
Average delivery time (rural)
2.3h
Operational range
100km
Maximum payload
100kg
Delivery success rate
98%
Road alternative (median)
3–5 days
Direction
What comes next.
Hardware
Smart textile biosensors — continuous monitoring embedded in the fabric of daily life rather than a discrete wearable. Expanded sensor suite targeting analytes that currently require lab draws: glucose approximation, hydration, electrolytes.
Ask Helios — direction
Today Ask Helios navigates the Helios platform. The direction is a genuine health companion: consolidating records across providers, surfacing research relevant to your condition, summarising consultation notes, and helping you understand your own health history in plain language. These capabilities move from direction to shipped as the underlying infrastructure is confirmed.
Software
Phase 2 moves to database-backed inference systems with direct access to longitudinal biometric records. Structured queries against a patient's full health history. Clinical decision support that integrates payment, care access, and health data in a single context window.
Geography
15+ countries across sub-Saharan Africa. Each deployment generates population-level data that improves model accuracy for subsequent deployments. The infrastructure compounds. We are building for 2.5 billion people. The products you see today are the first nodes in that network.
Security & Compliance
Medical data demands medical-grade security.
The platform handles biometric data, clinical records, and financial transactions simultaneously. The security architecture is designed from that constraint outward.
Encryption at rest and in transit
All health and personal data is encrypted at rest and in transit. Raw waveform data does not transit the network until a secure sync path is confirmed.
Role-based access control
Access to patient data is scoped by role. Patients, doctors, facility staff, community health workers, and administrators each have precisely bounded permissions. No role has broader access than its function requires.
Full audit logging
Every data access event is recorded: timestamp, accessor identity, and declared purpose. Audit logs are tamper-evident. If anyone accessed your record, you can see who, when, and why.
Consent-gated access
Healthcare providers request access. Patients grant it — for specific record types, for specific time windows. The default is no access. Emergency access exists for genuine emergencies: it is logged, time-limited, and the patient is notified.
Regulatory alignment
Built to comply with the Uganda Data Protection and Privacy Act (DPPA) 2019. Medical device application submitted to the Uganda National Drug Authority.
Get in Touch
Want to understand what we're building?
We talk to engineers, investors, health system operators, and researchers. If the problem is relevant to you, we want to hear from you.
Or write directly: centraloffice@helioshealths.com