"""Evaluate the locally generated checkpoint against the processed DDInter dataset. Produces: - MEDCARE-DDI-AI/models/eval/metrics_summary.json - MEDCARE-DDI-AI/models/eval/confusion_matrix.png - MEDCARE-DDI-AI/models/eval/inference_validation_report.md The script re-uses preprocessing logic from predictor.py to ensure consistency. """ from __future__ import annotations import json import math from collections import Counter from pathlib import Path from typing import Any import matplotlib.pyplot as plt import numpy as np import pandas as pd from preprocessing.artifact_manager import manager import torch from sklearn.metrics import accuracy_score, confusion_matrix, f1_score, recall_score from predictor import ( BASE_DIR, DATA_PATH, MODEL_PATH, LABEL_NAMES, LABEL_TO_INDEX, INDEX_TO_LABEL, normalize_name, canonical_pair_key, HybridDDIPredictor, ) OUT_DIR = MODEL_PATH.parent / 'eval' OUT_DIR.mkdir(parents=True, exist_ok=True) def build_pairs_from_csv(df: pd.DataFrame) -> pd.DataFrame: # For evaluation: collapse multiple evidence rows into one canonical pair pairs = {} for _, row in df.iterrows(): a = str(row['Drug_A']).strip() b = str(row['Drug_B']).strip() level = str(row['Level']).strip().lower() key = canonical_pair_key(a, b) if key not in pairs: pairs[key] = Counter() pairs[key][level] += 1 records = [] for (a, b), counter in pairs.items(): # majority label label, _ = counter.most_common(1)[0] records.append({'drug_a': a, 'drug_b': b, 'label': label, 'support': sum(counter.values())}) return pd.DataFrame(records) def evaluate(predictor: HybridDDIPredictor, eval_df: pd.DataFrame) -> dict[str, Any]: y_true = [] y_pred = [] oov_count = 0 for _, row in eval_df.iterrows(): a = row['drug_a'] b = row['drug_b'] label = row['label'] y_true.append(LABEL_TO_INDEX.get(label, 0)) # Use predictor internals to produce logits and probabilities a_id = predictor._find_vocab_id(a) b_id = predictor._find_vocab_id(b) if a_id == 0 or b_id == 0: oov_count += 1 with torch.no_grad(): logits = predictor.model(torch.tensor([a_id], dtype=torch.long), torch.tensor([b_id], dtype=torch.long)) probs = torch.softmax(logits, dim=-1).squeeze(0).cpu().numpy() pred_idx = int(np.argmax(probs).item()) y_pred.append(pred_idx) y_true = np.array(y_true) y_pred = np.array(y_pred) acc = float(accuracy_score(y_true, y_pred)) macro_f1 = float(f1_score(y_true, y_pred, average='macro', zero_division=0)) # severe recall corresponds to 'major' label if 'major' in LABEL_TO_INDEX: major_idx = LABEL_TO_INDEX['major'] severe_recall = float(recall_score(y_true, y_pred, labels=[major_idx], average='macro', zero_division=0)) else: severe_recall = 0.0 cm = confusion_matrix(y_true, y_pred, labels=list(range(len(predictor.label_names)))) metrics = { 'accuracy': round(acc, 4), 'macro_f1': round(macro_f1, 4), 'severe_recall': round(severe_recall, 4), 'num_examples': int(len(eval_df)), 'oov_count': int(oov_count), 'oov_rate': round(float(oov_count) / max(1, len(eval_df)), 4), } return metrics, cm, y_true, y_pred def save_confusion_matrix(cm: np.ndarray, labels: list[str], out_path: Path) -> None: fig, ax = plt.subplots(figsize=(6, 5)) im = ax.imshow(cm, interpolation='nearest', cmap=plt.cm.Blues) ax.figure.colorbar(im, ax=ax) ax.set(xticks=np.arange(cm.shape[1]), yticks=np.arange(cm.shape[0]), xticklabels=labels, yticklabels=labels, ylabel='True label', xlabel='Predicted label', title='Confusion Matrix') plt.setp(ax.get_xticklabels(), rotation=45, ha='right', rotation_mode='anchor') thresh = cm.max() / 2.0 for i in range(cm.shape[0]): for j in range(cm.shape[1]): ax.text(j, i, format(int(cm[i, j]), 'd'), ha='center', va='center', color='white' if cm[i, j] > thresh else 'black') fig.tight_layout() fig.savefig(out_path, dpi=150) plt.close(fig) def main() -> None: print('Loading checkpoint and predictor...') if not MODEL_PATH.exists(): raise FileNotFoundError(f'Checkpoint not found at {MODEL_PATH}') predictor = HybridDDIPredictor.from_default_paths() print('Loading processed dataset...') df = manager.load_artifact('ddinter_combined') eval_df = build_pairs_from_csv(df) print(f'Prepared {len(eval_df)} canonical pairs for evaluation') metrics, cm, y_true, y_pred = evaluate(predictor, eval_df) # Additional checks: preprocessing consistency metadata = { 'model_version': predictor.model_version, 'vocab_size_checkpoint': len(predictor.vocab), 'vocab_size_used_by_model': predictor.model.embedding.num_embeddings - 1, 'embedding_dim_checkpoint': predictor.embedding_dim, 'model_embedding_dim': predictor.model.embedding.embedding_dim, 'label_names': predictor.label_names, 'index_to_label': predictor.index_to_label, 'num_eval_pairs': len(eval_df), } metrics.update(metadata) # Save metrics JSON metrics_path = OUT_DIR / 'metrics_summary.json' with metrics_path.open('w', encoding='utf-8') as fh: json.dump(metrics, fh, indent=2) # Save confusion matrix PNG cm_path = OUT_DIR / 'confusion_matrix.png' save_confusion_matrix(cm, predictor.label_names, cm_path) # Generate simple report report_lines = [] report_lines.append('# Inference Validation Report') report_lines.append('') report_lines.append(f'- Model version: {predictor.model_version}') report_lines.append(f'- Eval pairs: {len(eval_df)}') report_lines.append(f"- Vocab size (checkpoint): {metadata['vocab_size_checkpoint']}") report_lines.append(f"- Vocab size (model): {metadata['vocab_size_used_by_model']}") report_lines.append(f"- Embedding dim (checkpoint): {metadata['embedding_dim_checkpoint']}") report_lines.append(f"- Embedding dim (model): {metadata['model_embedding_dim']}") report_lines.append('') report_lines.append('## Metrics') report_lines.append('') report_lines.append(f"- Accuracy: {metrics['accuracy']}") report_lines.append(f"- Macro F1: {metrics['macro_f1']}") report_lines.append(f"- Severe (major) recall: {metrics['severe_recall']}") report_lines.append(f"- OOV count: {metrics['oov_count']} (rate: {metrics['oov_rate']})") report_lines.append('') report_lines.append('## Confusion matrix') report_lines.append(f'Confusion matrix saved to `{cm_path}`') report_lines.append('') report_lines.append('## Preprocessing & Consistency Checks') report_lines.append('- Label ordering (checkpoint): ' + ', '.join(predictor.label_names)) report_lines.append('- Index to label mapping:') report_lines.append('') for idx, label in predictor.index_to_label.items(): report_lines.append(f'- {idx} -> {label}') report_lines.append('') report_lines.append('## Observations & drift checks') if metrics['oov_rate'] > 0.05: report_lines.append('- Warning: OOV rate exceeds 5% — incoming drug names differ from training vocabulary.') else: report_lines.append('- OOV rate within expected bounds.') # Healthcare grade quick pass/fail on severe recall threshold_severe_recall = 0.90 if metrics['severe_recall'] >= threshold_severe_recall: report_lines.append(f'- Severe recall >= {threshold_severe_recall} (PASS)') else: report_lines.append(f'- Severe recall < {threshold_severe_recall} (FAIL) — consider retraining or calibration for higher sensitivity on critical events') report_path = OUT_DIR / 'inference_validation_report.md' report_path.write_text('\n'.join(report_lines), encoding='utf-8') print('Evaluation complete.') print(f'Metrics JSON: {metrics_path}') print(f'Confusion matrix: {cm_path}') print(f'Report: {report_path}') if __name__ == '__main__': main()