Update README.md

README.md

@@ -89,3 +89,133 @@ probs = torch.softmax(logits, dim=-1).squeeze()
 p_ohca = float(probs[1])   # index 1 = OHCA, index 0 = Non-OHCA
 
 print({"p_ohca": p_ohca})
+
+Decision threshold
+
+This is a probability cutoff for calling a note “OHCA.” You can tune it to your setting:
+
+High sensitivity (screening): 0.28–0.32
+
+Balanced: 0.36 (v8 validation-optimized neighborhood)
+
+Higher precision: 0.50+
+
+def predict_ohca(text, threshold=0.32):
+    inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=512)
+    with torch.no_grad():
+        logits = model(**inputs).logits
+    p = torch.softmax(logits, dim=-1)[0,1].item()
+    label = "OHCA" if p >= threshold else "Non-OHCA"
+    return {"label": label, "p_ohca": p, "threshold": threshold}
+
+print(predict_ohca(text, threshold=0.32))
+
+Data and preprocessing
+
+Source: MIMIC-derived discharge notes (internal processing).
+
+Sections used:
+
+Chief Complaint
+
+History of Present Illness (also recognized as “History of Present Illness:” / “HPI”)
+
+Class distribution (330 total):
+
+Non-OHCA: 242 (73.3%)
+
+OHCA: 47 (14.2%)
+
+Inter-facility transfers: 23 (7.0%)
+
+In-hospital arrests: 18 (5.5%)
+
+Only the binary OHCA vs Non-OHCA head is used at inference here. Multi-class labels helped training signal.
+
+Splits (patient-level): Train 210, Val 54, Test 66 unique admissions.
+
+Max length: 512 tokens.
+
+Training
+
+Base: PubMedBERT (microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract)
+
+Epochs: 5–6 with continued improvement through epoch 4
+
+Batching: small batch + gradient accumulation
+
+Sampler: class-balanced mini-batches
+
+Loss: weighted cross-entropy (class imbalance)
+
+Optimizer/schedule: AdamW + linear decay
+
+Hardware: CPU training (works for inference on CPU)
+
+Evaluation (test set)
+
+Confusion matrix at a recall-oriented operating point:
+
+	Pred Non-OHCA	Pred OHCA
+Actual Non	51	7
+Actual OHCA	0	9
+
+Metrics
+
+Sensitivity (Recall): 1.000
+
+Specificity: 0.879
+
+Precision (PPV): 0.562
+
+NPV: 1.000
+
+F1-score: 0.720
+
+ROC-AUC: 0.971
+
+Interpretation: At this threshold, the model did not miss any OHCA cases in the test set, at the cost of 7 false positives.
+
+Threshold selection guide
+
+Pick a threshold that matches your use case:
+
+Screening (don’t miss OHCA): 0.28–0.32
+
+Balanced review load: around 0.36
+
+Fewer false positives: ≥ 0.50
+
+If you need to optimize explicitly for recall, compute the best Fβ with β>1 (e.g., F2) on your validation set and use that threshold in production.
+
+Limitations
+
+Trained on a specific documentation style; performance may vary on different systems.
+
+English only; text quality and section headers matter.
+
+Always keep a human-in-the-loop for high-stakes decisions.
+
+Citation
+
+If you use this model, please cite:
+
+M. Moukaddem. OHCA Classifier v8: PubMedBERT fine-tuned for Out-of-Hospital Cardiac Arrest detection in discharge notes. 2025. https://huggingface.co/monajm36/ohca-classifier-v8
+
+License
+
+MIT
+
+
+---
+
+## Optional polish (recommended)
+- On the model page, click **Edit model card** and fill the left-hand **Metadata UI**:
+  - `license`: mit
+  - `language`: en
+  - `base_model`: microsoft/BiomedNLP-PubMedBERT-base-uncased-abstract
+  - `pipeline_tag`: text-classification
+- Turn on the **Inference widget** so people can paste text and see probabilities.
+- Add a **README image** later (e.g., a small confusion matrix figure).
+
+If you want, I can also generate a small Python snippet that reads your validation set, computes **F2-optimal** threshold, and prints a compact threshold table to include in the card.