Research Prototype

Clinical AI that
explains itself.

ShifaMind predicts ICD-10 diagnoses from clinical notes while providing concept-grounded explanations clinicians can trust and verify. Explainability by design — not afterthought.

0.4727
Diagnostic F1 i Macro-averaged F1 across 50 ICD-10 codes with extreme class imbalance
115K
Clinical Notes
50
ICD-10 Codes
1.0
Completeness
How ShifaMind Works
Every prediction flows through human-interpretable clinical concepts. No shortcuts. No black boxes.
01

Text Encoding

BioClinicalBERT encodes discharge summaries into rich clinical representations

02

Concept Grounding

Cross-attention grounds 111 clinical concepts (symptoms, findings, treatments)

03

Multiplicative Gate

Concept activations multiply with diagnosis embeddings — zero concepts = zero signal

04

Diagnosis + Confidence

ICD-10 predictions with concept-level confidence scores for full transparency

Performance & Interpretability
Competitive diagnostic accuracy with full concept-mediated transparency — evaluated on MIMIC-IV.
Diagnostic F1 (Top-50 ICD-10)
Tuned Threshold
ShifaMind
0.4727
LAAT
0.4637
Vanilla CBM
0.4398
CAML
0.4287
MSMN
0.4165
GPT-5.4
0.4122
Claude 4.6
0.3556
Interpretability Metrics
Enforced
1.000
Completeness
0.904
TCAV Score
0.744
Faithfulness
Concept Completeness = 1.0 means every prediction is fully mediated by clinical concepts. The multiplicative bottleneck mathematically guarantees: if concepts are zero, the diagnosis signal is zero.
Sample Prediction Output
A simulated ShifaMind inference on a clinical discharge summary.
shifamind_inference.py — Patient #48201
Input: Discharge Summary (1,247 tokens)
72M admitted with progressive dyspnea, bilateral lower extremity edema, and elevated BNP. CXR showing bilateral pleural effusions. Started on IV furosemide with improvement...
Activated Concepts:
edema
0.94
diuretics
0.91
cardiac
0.88
dyspnea
0.86
pleural_effusion
0.79
Prediction:
[1] I50.9 — Heart failure, unspecified conf: 0.92
[2] J91.8 — Pleural effusion conf: 0.78
[3] I10   — Essential hypertension conf: 0.71
Concept Confidence Scores:
I50.9: edema(0.94) × cardiac(0.88) × diuretics(0.91) → 0.92
J91.8: pleural_effusion(0.79) × dyspnea(0.86) → 0.78
Simulated output for illustration purposes only. Predictions based on de-identified MIMIC-IV data.

See ShifaMind in Action

Doctors can analyze real clinical notes, explore concept activations, and discuss cases with AI — all grounded in ShifaMind’s interpretable predictions.

Streaming AI Chat Real-time Predictions Doctor Feedback
Access Platform → Limited access · Contact us for a demo account.
Future Vision
Building toward the future of transparent clinical decision support.
Multi-Agent Specialist Network
Dedicated AI agents each specializing in a disease category, orchestrated by a central routing agent for domain-expert accuracy and full explainability.

Cardiology Agent

Deep specialization in heart failure, arrhythmias, and vascular conditions with cardiology-specific concept vocabularies

Pulmonary Agent

Expert-level reasoning over respiratory conditions — COPD, pneumonia, pleural diseases

Renal Agent

Specialized in kidney pathology with lab-value-aware concept grounding and AKI staging

Orchestrator Agent

Routes clinical cases to the right specialist, aggregates multi-domain findings, resolves conflicting signals

Why ShifaMind
Transparent clinical AI built for the trust requirements of healthcare.

Interpretability Gap

Current clinical AI is black-box. ShifaMind is the first system with mathematically guaranteed concept completeness (1.0) — every prediction is fully mediated by human-interpretable clinical concepts.

Market Opportunity

Clinical NLP and medical coding automation is a multi-billion dollar market. Hospital systems spend billions on manual coding. Interpretable AI that clinicians trust is the unlock for widespread adoption.

Technical Moat

Novel architecture: multiplicative concept bottleneck with cross-attention fusion. Outperforms vanilla concept models by 7.5% while maintaining perfect interpretability. Published methodology, defensible IP.

Roadmap

Phase 1: Core model — complete
Phase 2: Knowledge graph + RAG evidence integration
Phase 3: Upgrade base encoder to BioClinicalModernBERT
Phase 4: Multi-agent specialist network
About
Mohammed Sameer Syed
Founder · AI/ML Engineer
Building ShifaMind at the intersection of clinical AI and explainability. Focused on creating systems where every diagnostic prediction is transparent, verifiable, and grounded in medical knowledge — because clinicians shouldn’t have to trust a black box.
🎓 M.S. Machine Learning 🏥 MIMIC-IV · BioClinicalBERT · UMLS