Research-stage  ·  Open for collaboration

Helping cardiologists catch heart failure before it becomes a readmission.

A cardiac AI pipeline built on real patient data across 6 hospitals in 4 countries. No synthetic data. No corporate funding.

30,000+Real patient records
6Hospitals
4Countries

LangGraph · Gemini · XGBoost · SHAP · EchoNet · MONAI · FastAPI

Research-stage MVP  ·  Not a cleared medical device

01 — The clinical gap

Why this pipeline needs to exist.

15.8%

of stent patients are readmitted within 30 days of the procedure.

Most readmissions happen because warning signs in biomarker trends go unnoticed between discharge and the first follow-up.

JAMA Cardiology
25%

of post-PCI patients return to hospital within 12 months.

There is no continuous monitoring layer between cardiology appointments. Each visit sees only the latest report — not the trend.

Lancet Digital Health
2.4%

early mortality within 30 days post-procedure — rising to 10% at 5 years.

Delayed intervention on recurring blockages or early heart failure signs is the leading cause of this gap.

European Heart Journal
Hours

lost per patient when cardiologists manually correlate ECG, DICOM, and blood panel data.

Cognitive overload from unstructured multi-modal data leads to missed early warning patterns.

Journal of the American College of Cardiology

Keep scrolling — each stat is a published outcome, not a projection

02 — The pipeline

How the intelligence pipeline works.

Deterministic ML handles the risk scoring. Gemini only extracts and polishes language — never interprets numbers.

01

Upload any cardiac document

Live now

01 / 06

Upload any cardiac document

PDF reports, ECG printouts, blood panel results, or CT Angiogram reports. The system accepts messy, real-world hospital documents — not clean structured data.

Post-procedure PDFEcho reportBlood panelCT Angiogram

02 / 06

LangGraph + Gemini extracts structured values

A multi-node LangGraph pipeline detects the report type, then runs a cardiac-specific Gemini prompt to extract the exact biomarkers cardiologists track. Unstructured text becomes structured JSON.

Detects: Echo / CT Angio / Blood PanelExtracts 9 cardiac biomarkersReturns confidence score per field

03 / 06

XGBoost ensemble scores tabular risk

Three specialized XGBoost sub-models run in parallel — one for Acute MI risk, one for Structural Failure, one for MACE (Major Adverse Cardiac Events). Their outputs are weighted into a single risk trajectory.

Sub-model 1: Acute MISub-model 2: Structural FailureSub-model 3: MACE RiskEnsemble: weighted average

04 / 06

EchoNet-Dynamic reads Echo videos for EF%

A pre-trained deep learning model (Stanford open-source) processes echocardiogram video frames to predict Ejection Fraction — the most critical post-stent cardiac metric.

Echo video input (DICOM)EF% prediction from video framesMore accurate than text extraction alone

05 / 06

MONAI analyzes CT Angiogram for stenosis

MONAI framework processes CT Angiogram images to detect and quantify coronary artery stenosis percentage and calcium score — directly from the imaging data, not just the radiology text report.

CT Angiogram DICOM inputLAD stenosis % from imagesCalcium score quantification

06 / 06

SHAP explains exactly why a flag was raised

Every risk signal is backed by SHAP (SHapley Additive Explanations). The system shows which biomarkers drove the output — so cardiologists see the reasoning, not a black-box score.

Feature importance per biomarkerHardcoded clinical templates per markerPlain language via Gemini polish

03 — Models we use

Multi-modal from the ground up.

Each data type needs a different model. We are honest about what is built today and what is coming next.

01

Tabular Data

XGBoost Ensemble

Live — Phase 1

Three specialized gradient-boosted tree models trained exclusively on real patient records. No LLM involvement in risk scoring — purely deterministic ML.

  • Sub-model 1: Acute MI detection
  • Sub-model 2: Structural failure risk
  • Sub-model 3: MACE risk scoring
  • Biomarker velocity features (NT-proBNP trend)
  • SHAP explainability on every output
02

ECG & Echo

EchoNet-Dynamic

Integrating — Phase 2

Stanford open-source deep learning model trained on 10,000+ echocardiogram videos to predict Ejection Fraction directly from the video — more accurate than text extraction alone.

  • Input: Echo video (DICOM format)
  • Output: EF% predicted from video frames
  • Trained by Stanford ML Group
  • Open-source, no licensing cost
03

CT Angiogram

MONAI Framework

Planned — Phase 2

MONAI (Medical Open Network for AI) is the standard framework for medical imaging AI. Used here to process CT Angiogram DICOM files and quantify coronary stenosis and calcium score.

  • Input: CT Angiogram DICOM
  • Calcium score quantification
  • LAD stenosis % from image data
  • PyTorch-based, production-ready

04 — Data foundation

Built on real-world evidence. Not synthetic assumptions.

Healthcare AI trained on synthetic data fails in real clinical environments. Every record here came from a real patient in a real hospital.

0+

Real patient records

0

Hospitals

0

Countries

Coronary CTA Risk Reclassification Cohort

Structural imaging · Calcium score · LAD stenosis · MACE risk

27,126

Netherlands / USA

Zheen Hospital, Erbil

CK-MB · Troponin · Acute MI · Real clinical pipeline

1,319

Iraq

Zia Heart Foundation Hospital

South Asian cardiac phenotypes · Troponin-I · LDL · eGFR

1,048

Bangladesh

UCI Multi-Hospital Consortium

Cleveland Clinic · Gold-standard benchmark · 4-center validation

920

USA / Hungary / Switzerland

University Medical Center — HF Clinical Records

EF% · Serum creatinine · Heart failure outcomes · Real patients

299

Netherlands

NT-proBNP Longitudinal HFrEF Cohort

2-timepoint NT-proBNP · Velocity features · Mortality outcomes

157

Netherlands

Zero synthetic data  ·  Zero SMOTE augmentation

Live demo

See it work on a real cardiac report.

Upload any cardiac PDF — an Echo report, CT Angiogram, blood panel, or prescription. The pipeline extracts cardiac biomarkers and returns structured values in seconds.

No account required. No data stored. Your file is processed and discarded immediately.

Report type detectedEF% extractedLAD stenosis %Troponin-I valueCreatinine + eGFRLDL CholesterolCalcium ScoreConfidence score per field
Upload a report and try it →

No cardiac report? The demo page includes sample test data to try.

05 — Who I am

My name is Avinash.
I work in data science and applied AI.

I build applied AI systems — retrieval-augmented generation, agentic pipelines with LangChain and LangGraph, and machine learning models that go from raw data to a production API. OjasLabs is where I applied those skills to a real clinical problem.

The LangGraph extraction pipeline, the XGBoost risk models, the FastAPI backend, the Next.js frontend — I built all of it end to end, on 30,000+ real patient records, to prove I can take an applied AI system from idea to something a hospital could use.

I want to go deeper into this kind of work — data science, applied AI engineering, RAG and LLM systems — at a company building in health-tech, clinical AI, or applied ML. If that's you, I'd like to talk.

What I am looking for:

Hire Me

Data Scientist · Applied AI Engineer · RAG / LLM / LangChain

Hands-on with retrieval-augmented generation, agentic pipelines built on LangChain and LangGraph, and machine learning from raw data to production APIs. Looking for a role that goes deeper into applied AI — ideally in health-tech or clinical AI.

View what I built

Co-Found

Clinical · Regulatory · Business

The technical foundation is here. To take this to hospitals requires clinical validation, regulatory navigation (MDR, FDA), and business development. If that is your background, let's talk.

Get in touch

Research Partner

Cardiologists · Hospitals · Academia

I am looking for a cardiology department willing to validate this pipeline on retrospective patient data. In return: early access to the platform, custom integration, and if it leads somewhere worth publishing, I'd welcome working on that together.

Start a conversation