Clinical-Emotisupport/NLP-Clinical-Emotisupport

Clinical EmoTiSupport v2 (Business-Aligned Logic) 🏥

🚀 Overview: Beyond Standard Metrics

Clinical EmoTiSupport v2 is not just a text classifier; it is a risk-calibrated engine designed for real-world digital health environments.

Unlike standard NLP models that treat all errors equally, this model was fine-tuned and calibrated with a specific business philosophy: In healthcare, a False Negative on 'Anger' or 'Anxiety' is more dangerous than a False Positive.

We refined the dataset to remove logical inconsistencies (e.g., "polite anger") and implemented a Safety Guardrail mechanism to filter out neutral administrative noise with 94% precision.

💡 Key Upgrades in v2

1. Cleaner Data (v6): We aggressively cleaned the training data (2,000 samples) to remove "Politeness Artifacts" — specifically cases where patients used "Thank you" while expressing severe frustration, which confused previous iterations.
2. Adaptive Thresholding: We moved away from the standard 0.5 threshold. This model is designed to be used with class-specific thresholds (see Inference Logic below) to maximize the detection of urgent signals.
3. Neutral Guardrail: A dedicated logic layer that prevents low-confidence predictions from triggering false alarms.

📊 Performance & Calibration

Evaluated on a strictly cleaned validation set (20% split), calibrated for high sensitivity on critical emotions:

Metric	Value	significance
Micro F1-Score	0.84	Robust overall performance
Neutral Precision	0.94	Highly trustworthy when it says "Nothing is wrong"
Anger Recall	~0.60	+23% improvement over v1 (Critical for retention)

Recommended Thresholds (Business Logic)

To reproduce our results, do not use a flat 0.5 cutoff. Use the following sensitivity thresholds:

Emotion	Threshold	Reasoning
Anger	0.30	High Sensitivity: Better to flag a false alarm than miss a furious patient.
Anxiety	0.40	High Sensitivity: Early detection of distress signals.
Frustration	0.40	Early warning sign of churn.
Disappointment	0.50	Balanced precision/recall.
Confusion	0.50	Balanced precision/recall.
Satisfaction	0.65	High Precision: Only classify as "Satisfied" if certainty is high (avoids "polite" false positives).
Guiderail	0.35	If max(probs) < 0.35, the label is Neutral.

🛠 How to Use (with Business Logic)

To get the best results, use this inference snippet which includes the Guardrail and Adaptive Thresholds:

import torch
import numpy as np
from transformers import AutoTokenizer, AutoModelForSequenceClassification

# 1. Setup
model_id = "YourUserName/Clinical-EmoTiSupport-v2" # Update this!
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)

# 2. Define Business Logic
THRESHOLDS = {
    "anxiety": 0.40, "confusion": 0.50, "frustration": 0.40, 
    "anger": 0.30, "disappointment": 0.50, "satisfaction": 0.65
}
EMOTIONS = ["anxiety", "confusion", "frustration", "anger", "disappointment", "satisfaction"]
GUARDRAIL = 0.35

# 3. Predict
text = "I've been waiting for my results for a week, this is ridiculous!"
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    probs = torch.sigmoid(model(**inputs).logits).squeeze()

# 4. Apply Logic
results = {}
if probs.max() < GUARDRAIL:
    print("Prediction: Neutral / Administrative")
else:
    for i, label in enumerate(EMOTIONS):
        if probs[i] > THRESHOLDS[label]:
            results[label] = float(probs[i])

print("Detected Emotions:", results)
# Expected Output: {'anger': 0.72, 'frustration': 0.81} (Example)