AI
Patient devices, manual entries, and physician oversight unified in one record, with machine learning-based insulin dose prediction. The secure cloud platform shows patients, caregivers, and physicians the data that matches their permissions. A Random Forest model learns from each patient's own history. The platform reduced four separate systems to a single, continuous, GDPR-compliant record.
Key Takeaways
One unified record instead of 4+ separate apps (4× fewer systems)
Personalized insulin dose prediction with a per-patient Random Forest model
A continuous, real-time record for authorized physicians
Three user roles (patient, caregiver, physician) with a full audit trail
Estimated savings of ≈ 10–15 hours per month per patient, with a ≈ 2–3x return within the first year
Background
Starting point
Apps and notes per patient: 4+ disconnected sources (device apps, food diaries, paper notes; nothing is synced)
Physician visibility between visits: None (only a snapshot of the patient's current state during the appointment)
Insulin dosing support: Manual estimation only (patients estimate doses from experience, with no data-driven support)
Access control for health data: None (no defined roles and no audit trail)
The challenge
Diabetes data fragmented across devices and apps - people with diabetes use continuous glucose monitors, glucose meters, food diaries, and paper notes every day, and none of it is connected. Physicians see only a snapshot during appointments, and insulin dosing depends entirely on the patient's own judgment.
Glucose readings are stored in closed manufacturer apps, with no shared view
Manual entries (meals, insulin, activity) are often kept in separate notes
Physicians have no continuous patient record between visits
Insulin doses are estimated by eye, with no data-driven support
Sensitive health data demands strict access control and an audit trail
Area | Before (without the platform) | After (with TAR's custom AI platform) |
|---|---|---|
Patient data sources | 4+ disconnected device apps, food diaries and paper notes | 1 unified cloud patient record, 4× fewer systems |
Data sync | Manual logging after the fact, nothing synced | Real-time sync of device readings and manual entries within seconds |
Insulin dosing | Manual dose estimate based on patient experience | Per-patient machine learning dose prediction (Random Forest model) |
AI personalization | No data-driven support for dosing decisions | ML model trained and tuned on each patient's own glucose, meal and activity history |
Physician visibility | Snapshot of the patient's state during appointments only | Continuous real-time record available between visits |
Health data privacy | Ad hoc storage and sharing of sensitive data | Encrypted storage and GDPR-compliant data handling by design |
Market size
≈ 540 million worldwide: people living with diabetes.
Estimated budget (NDA)
EUR 50,000–199,999
Budget breakdown:
Platform, patient record, and real-time sync: 40%
Machine learning model for insulin dose prediction: 25%
Security, GDPR, and access control: 20%
Clinical beta and production launch: 15%
Delivery in four phases
Phase 1: Research, clinical process analysis, security foundations
Phase 2: Platform architecture, device integration specification, role model
Phase 3: Central patient record, real-time sync, insulin dose prediction model
Phase 4: Beta with clinical partners, audit, and production launch
Project
A single patient record with insulin dose prediction
We built a secure cloud platform that ingests device readings and manual entries and shows patients, caregivers, and physicians the data that matches their permissions. On top of the unified record, we added a machine learning model that predicts the required insulin dose from each patient's glucose, meal, and activity history.
Real-time sync: Device readings and manual entries are available to all authorized users within seconds.
ML-based insulin dose prediction: A Random Forest regression model, trained and tuned separately for each patient, predicts the required insulin dose as soon as enough diary entries have been collected.
Role-based access: Precisely defined permissions separate the patient, caregiver, and physician views, backed by a full audit trail.
Privacy by design: Encrypted storage, GDPR-compliant data handling, and explicit patient consent.
How the service changed for patients and physicians
The capabilities were reshaped end to end: glucose data became a continuous, shared record instead of scattered snapshots, and dosing gained data-driven support.
Where the data lives: Before: locked in closed manufacturer apps; After: a unified cloud record, accessible from any device. One shared source of truth for patient, caregiver, and physician.
Insulin dosing: Before: the patient's manual estimate; After: per-patient machine learning prediction. The model learns from the patient's own data history instead of relying on rules of thumb.
Physician visibility: Before: a snapshot during appointments; After: a continuous record over time. Decisions are based on trends, not individual readings.
Privacy protection: Before: ad hoc storage and sharing; After: encryption and GDPR compliance from day one. Health data is handled to the standard that regulations require.
Stack & services
Node.js / Express, TypeScript, PostgreSQL, real-time WebSocket, machine learning (Python, scikit-learn), cloud infrastructure, GDPR / security compliance.
Results
From scattered notes to a data product with predictive value (Investment → product impact, NDA)
Ratio: ≈ 2–3x - estimated platform value relative to labor cost, first year
Incremental revenue: ≈ 10–15 hours per month - estimated time saved per patient on manual logging and compiling data after the fact
Investment: EUR 50,000–199,999
Payback: Within the first year of operation
Method: Comparison of the time spent collecting and compiling patient data before and after rollout, for patients and physicians combined
Confidence: Medium - based on usage during the beta period; no claims are made about clinical outcomes
The figures describe process efficiency only; medical outcomes are the responsibility of the clinical partners.
One record instead of four separate systems
The platform merged glucose readings, food diaries, insulin tracking, and activity data from separate apps into a single continuous record available to physicians, with a machine learning model at its core that predicts the required insulin dose for each patient.
Patient data sources: Before: 4+ separate apps; After: 1 unified view. 4× fewer systems (at platform launch).
Insulin dose prediction: Before: the patient's manual estimate; After: a per-patient machine learning model. Personalized — Random Forest regression, activated after ≈ 1,100 diary entries; the model is trained and tuned separately for each patient.
Physician visibility: Before: a snapshot at appointments; After: a continuous real-time record. Always available (to authorized physicians).
Supported user roles: Before: no structured access model; After: Patient · Caregiver · Physician. 3 role types (with a full audit trail).
How long does it take to develop a custom healthcare software platform?
What are the benefits of a unified patient record for physicians?
How is sensitive health data protected in a healthcare app?
How does machine learning predict insulin doses?
What is a diabetes management platform with machine learning?
Can't find the question you are looking for?
Send us a message and we'll get back to you as soon as possible!
Tell us where your AI is stuck.
An idea, a stalled pilot, or a phone that won't stop ringing unanswered – the first call is a working session, not a pitch.
