remote patient monitoring | Notion

https://youtu.be/ImkiFH3pj2Y

1. reality check – will this actually help the clinician and the patient?

quick questions	why they matter
what’s the exact clinical decision or action? e.g., “titrate meds if SBP > 160 mmHg for 3 readings.”	no decision ⇒ no value
minimum signal set to support that call?	don’t add sensors “just in case”
who pays? medicare RPM codes 99453/99454 in the US , NHIF or grant in Kenya	no payer ⇒ no adoption
regulatory lane? FDA non-device CDS vs. SaMD	skipping this comes back to bite
operational choke-point? network gaps, alert fatigue, patient compliance	fix these early

2. reference architecture (keep it boring)

2.1 capture

reusable ECG/multivital patch – 128 Hz ECG, skin-temp, accel, 30-day cache
add SpO₂ cuff or BP cuff only if it changes a treatment decision—specs are similar

2.2 pipe

BLE → phone → MQTT/HTTPS → S3 or a FHIR store.
Map raw packets to the HL7 Personal Health Device FHIR IG to stay vendor-agnostic .

2.3 compute

start with a rule engine (“3 out-of-range BP readings fire an SMS”), then iterate with ML.
always log why the alert fired to fight alert fatigue.

3. open datasets & simulators (click to download)

dataset / tool	what you get	link
PTB-XL – 21 k 12-lead ECGs	arrhythmia benchmarks	physionet
MIMIC-III Waveform DB	3 M h of ICU waveforms	physionet
VitalDB	6 388 surgeries @ 500 Hz	vitaldb.net
PulseDB	14 570 h ECG-PPG-ABP segments	frontiersin
Synthea	synthetic FHIR/CSV EMRs	GitHub
Open mHealth schemas	JSON specs for HR, BP, steps…	Open mHealth
FHIR PHD IG examples	ready-made Observation payloads	HL7 build site

4. common pitfalls