OtoskopiAI — Presentation Mode

OtoskopiAI

Smart ENT Screening Platform — Medinovatech

99.7%

V1 Best Val Acc

test 99.0%

100%

V2 Best Val Acc

test 99.7%

Sonnet 4.6

Claude AI Gate

OOD 3-lapis

dr. Yanuar Iman Santosa, Sp.T.H.T.B.K.L., Subsp.A.I. (K), M.Si.Med · Juni 2026 · v5.0

Problem & Solution

10.000+

Faskes primer Indonesia

vs hanya 1.827 Sp.THT

OMSK prevalensi 3,9% — gap diagnostik besar

Input: Foto otoskopi + anamnesis
Output: Diagnosis + confidence + rujukan
AI: MobileNetV2 + Claude Gate + Grad-CAM
Privacy: Klasifikasi self-hosted, Claude = cloud

Tech Stack

Frontend

HTML/CSS/JS + GSAP

Backend

Python 3.12 + FastAPI

AI/ML

PyTorch + TorchVision

Model CV

MobileNetV2 (ImageNet)

LLM Gate

Claude Sonnet 4.6 CLI

Training

FixMatch + SupCon Loss

Self-hosted VPS 16GB · Nginx + systemd · SQLite · Grad-CAM XAI

Arsitektur Sistem

Foto masuk → MobileNetV2 klasifikasi → OOD gate 3-lapis (entropy + energy + Claude) → diagnosis + rujukan PNPK OMSK 2018

Model Performance

99.7%

V1 Best Val

image-only · 3.4M params

100%

V2 Best Val

multimodal · 29-dim anamnesis

2.209

Training Images

5 kelas · balanced <19% gap

V2 val loss 0.0117 — decision boundary 3.8x lebih kuat dari V1 (0.0449)

Training Journey

Dataset > arsitektur. Lompatan akurasi = data bertambah, bukan model berubah.

26 Mei

91.7%

~73 gambar

28 Mei

91.7%

~200 gambar — masih mentok

30 Mei

99.4%

3.038 gambar — konvergen

Teknik Training

Transfer Learning

MobileNetV2 ImageNet → fine-tune untuk otoskopi

FixMatch SSL

Pseudo-labeling + augmentasi otomatis

SupCon Loss

Intra-class similarity, inter-class dissimilarity

99.7%

V1 · val loss 0.0449

100%

V2 · val loss 0.0117

V1 vs V2

F1-score per kelas — Sumotosima dataset

V1: Image-only, 1-2 detik, quick triage
V2: + 29-dim anamnesis, edge case handling
Claude Gate: OOD rejection + PNPK guidance
Keduanya saling melengkapi dalam produksi

Fitur & Timeline

Screening AI 2-3 detik (dual-panel V1+V2)
Claude Gate: OOD + interpretasi klinis PNPK
Urgency triage 3-level
Grad-CAM heatmap (XAI)
Human-in-the-loop retraining
Dashboard analytics + labeling

Continuous learning loop

External Validation — Balanced

Sumotosima dataset: 500 gambar, 100/kelas, sumber independen

99.4%

V1 Accuracy

98.2%

V2 Accuracy

>0.999

Macro AUC

CM V1 — 99.4%

CM V2 — 98.2%

Domain Shift — Kejujuran Ilmiah

Data dari sumber sangat berbeda → akurasi anjlok. Ini dilaporkan transparan.

CM Heterogeneous — 13.5%

OOD: 100% diprediksi OMA

13.5%

Acc (domain shift)

0.74

Macro AUC (ranking masih ada)

Inilah alasan Claude Gate (OOD 3-lapis) sangat penting sebagai safety net

Roadmap

Prototipe → validasi klinis prospektif → multi-center trial → SaMD

Estimasi Nilai & Biaya

Development

6.967

Lines of Code

~50+ jam · 30+ commits

Rp 40 – 60 juta

Nilai setara proyek

Operasional / Bulan

~Rp 50K

Total biaya bulanan

Cloud AI inference: Rp 0
Database (SQLite): Rp 0
Claude Gate: via subscription