Science & Technology

Every heartbeat
is a fingerprint.

Cardiac screening in India isn't gated by intent. It's gated by physics, by training, by trust. SamaHealth attacks all three — from the sensor to the model to the audit ledger.

Signal physicsOn-device MLClinical evalDPDPA auditField hardware

The problem at scale

The detection gap is the largest in primary care, period.

India has the workers — over a million ASHAs — and the political will. What's missing is a tool that respects the field: sunlight, noise, low literacy, and a worker who has 240 visits a month and no time for a 10-step UI.

1.7M

Indians die from CVD each year

11×

rural-vs-urban cardiologist gap

21%

of strokes are preventable AFib

< 1%

of ASHAs have screening tools

Translation

From a beating heart
to digital information.

SamaBeat captures three signals at the chest — phonocardiogram, single-lead ECG, accelerometer — and fuses them in 800 milliseconds. The model doesn't see a strip. It sees a 12-dimensional time-series of cardiac state.

PCGHeart sound · S1/S2 · murmurs

ECGSingle-lead · rhythm · rate · QRS morphology

ACCContact pressure · respiration · posture

Our technology

Five layers. One screening event.

Every layer is owned by SamaHealth. Vendor lock-in is a clinical-safety risk we won't take.

Hardware

SamaBeat

PCG + ECG + accelerometer in one ECDH-paired puck. Designed in-house with an ISO 13485 contract manufacturer in Bengaluru.

BLE 5.2 ECDH handshake
48-hour battery
IP54 sealed

Signal

Wavelet denoise + SQI

Tropical-clinic noise floors are ugly. Adaptive wavelet denoising + signal-quality gates reject unusable clips before inference.

35 dB(A) ambient budget
Cycle alignment
Pre-inference reject

Model

On-device TFLite

Quantized rhythm + abnormality models. Sub-800 ms inference on a sub-₹10k Android. Ships only after 78 clinical paths pass.

INT8 quantization
Rhythm + AFib + flutter
Voice-flow Hindi/Tamil/Bengali/Marathi

Evaluation

Pre-registered with CTRI

Retrospective (n>5k) + prospective (n>400) + adversarial-rural gates. A drop in any gate triggers automatic block.

Three CTRI-registered studies
Stop-rules public
No endpoint retrofitting

Trust

DPDPA · asia-south1

Tink AEAD per-field via Android Keystore. Append-only audit on every PHI touch. Mumbai-only residency.

Field-level encryption
PHI scrubber on crashes
Reason-gated reveals

Clinical validation

Validated where it has to be used.
Not where it's easy.

Prospective · Biharn = 412

Sadar District Hospital, Begusarai

AFib sensitivity 91.4% vs 12-lead reference
Specificity 96.1%
Median screening time 3.4 min

Retrospective · Karnatakan = 5,127

St. John's, Bengaluru

India-only ECG corpus
F1 on AFib 0.89, on flutter 0.84
False-positive rate halved vs baseline

Adversarial · Tamil Nadun = 96

GovMC + Sama labs

Wet skin · fan noise · crying infant overlay
SQI gate caught 87% of unusable clips
Pre-inference, before model sees signal

Pre-registration

All three studies are pre-registered with CTRI. Protocols, primary endpoints, and stop-rules are public before enrollment. We will not retrofit endpoints.

Most reviewed