Paediatric nephrotic syndrome affects roughly 1.5 lakh children in India. Most never see a nephrologist. Most never get a dipstick read by a trained eye. KidneyCare puts that capability in the palm of a community health worker — at AIIMS-validated accuracy.
The problem
A urine dipstick is cheap. A clinician reading it correctly is expensive and rare. For paediatric nephrotic syndrome, frequent dipstick monitoring matters — recurring proteinuria spikes are how relapses get caught early. But in PHC/CHC settings in India, the qualified eye isn't always there.
Existing computer vision approaches don't scale: lighting varies, dipstick brands vary, image quality varies, and getting the diagnosis wrong on a child is unacceptable.
The system we built
Three-stage CNN pipeline. Stage 1 runs on-device on Android or iOS — object detection on the dipstick, image quality gate (rejects blurry uploads), 5000K LED light box protocol for consistent lighting. Stage 2 segmentation runs in secure cloud — segments each colour pad, normalises against ambient. Stage 3 classification produces grades (NEG / 1+ / 2+ / 3+ / 4+) with confidence scoring.
The safety floor is the most important design choice. Anything under 70% confidence outputs 'Indeterminate' — never a guess on a child's diagnosis. That's a deliberate design choice validated with the AIIMS Patna paediatrics team.
Privacy is built in. Each patient gets a 6-digit anonymisation PIN. The PIN-to-identity registry is in a separate store accessible only to the Principal Investigator. Raw urine images never leave the encrypted pipeline. Data stays in ap-south-1 (Mumbai). 5-year retention per ICMR 2017 guidelines.
The outcomes
On the pilot cohort at AIIMS Patna (n=30), 90%+ accuracy vs lab. Cost per screening reduced by 70–80% compared with a clinician-read dipstick at a tertiary centre. TRL moved from 4 to 6 across the pilot. CDSCO Class B regulatory pathway mapped. APIIC Ignition Grant + NIDHI-DST funded the validation.
"The Indeterminate output is the most important feature. Tell me you don't know rather than tell me something wrong about a child."
— paediatrics team, AIIMS Patna
↓ tls 1.3
stage 1 detection (on-device) → stage 2 segmentation → stage 3 classification
↓
confidence floor · pin anonymisation
↓
redcap crf · ap-south-1 only · icmr 5y retention
What's underneath
- Aminobots Lens — KidneyCare (the three-stage CNN pipeline)
- Mobile app (React Native · TensorFlow Lite on-device)
- AWS SageMaker for stages 2 & 3 (ap-south-1)
- Per-tenant CMK + private endpoints + India residency
- REDCap clinical CRF integration
- Blueprint System (AWS Inferentia target)