File size: 17,843 Bytes
6dc9d46
 
 
3ca1d38
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9659593
3ca1d38
 
6dc9d46
 
ad2e847
696f787
 
 
6dc9d46
696f787
 
6dc9d46
 
3ca1d38
 
 
6dc9d46
 
 
9659593
6dc9d46
 
 
 
 
 
9659593
6dc9d46
 
 
 
 
696f787
 
6dc9d46
 
 
 
 
 
9659593
6dc9d46
696f787
6dc9d46
 
aefac4f
 
6dc9d46
 
 
9659593
6dc9d46
 
 
9659593
3ca1d38
6dc9d46
3ca1d38
 
 
696f787
6dc9d46
9659593
696f787
9659593
 
 
 
 
6dc9d46
 
 
 
 
 
 
 
 
 
 
9659593
 
 
 
 
6dc9d46
 
 
 
 
 
 
 
 
 
 
 
9659593
 
 
 
696f787
6dc9d46
9659593
 
 
 
 
 
 
 
696f787
6dc9d46
 
 
 
9659593
aefac4f
ad2e847
 
6dc9d46
 
 
9659593
6dc9d46
 
 
 
 
9659593
6dc9d46
696f787
6dc9d46
9659593
 
696f787
6dc9d46
 
 
 
 
696f787
6dc9d46
 
696f787
6dc9d46
 
 
 
 
696f787
6dc9d46
 
 
 
9659593
6dc9d46
 
 
9659593
3ca1d38
6dc9d46
3ca1d38
 
 
696f787
6dc9d46
9659593
696f787
9659593
 
 
 
 
6dc9d46
 
 
 
 
 
 
 
 
 
 
 
9659593
 
 
 
 
6dc9d46
 
 
 
 
 
 
 
 
 
 
 
9659593
 
 
 
696f787
6dc9d46
9659593
 
 
 
 
 
 
 
 
696f787
6dc9d46
 
 
9659593
aefac4f
ad2e847
 
6dc9d46
 
 
9659593
6dc9d46
 
 
9659593
3ca1d38
6dc9d46
3ca1d38
 
 
696f787
6dc9d46
 
9659593
6dc9d46
 
 
696f787
6dc9d46
9659593
696f787
6dc9d46
 
 
 
 
 
 
 
9659593
6dc9d46
 
 
 
 
 
 
 
696f787
6dc9d46
 
9659593
 
 
 
 
 
 
6dc9d46
 
696f787
6dc9d46
 
 
696f787
6dc9d46
 
 
696f787
9659593
696f787
6dc9d46
 
 
 
 
 
 
 
 
 
 
 
 
 
696f787
6dc9d46
 
 
 
9659593
6dc9d46
 
 
 
 
696f787
6dc9d46
 
696f787
6dc9d46
 
 
696f787
6dc9d46
 
9659593
6dc9d46
9659593
6dc9d46
696f787
6dc9d46
9659593
6dc9d46
9659593
696f787
6dc9d46
 
696f787
6dc9d46
9659593
696f787
6dc9d46
9659593
6dc9d46
696f787
6dc9d46
 
aefac4f
6dc9d46
 
696f787
9659593
 
 
696f787
6dc9d46
 
 
 
 
 
 
 
 
696f787
6dc9d46
 
 
 
 
9659593
6dc9d46
 
 
 
 
 
 
696f787
6dc9d46
 
 
 
aefac4f
 
9659593
aefac4f
 
 
 
6dc9d46
696f787
6dc9d46
 
 
696f787
6dc9d46
 
696f787
6dc9d46
 
696f787
6dc9d46
 
696f787
6dc9d46
 
696f787
6dc9d46
 
 
696f787
6dc9d46
 
 
 
 
9659593
6dc9d46
3ca1d38
 
 
 
 
 
9659593
 
3ca1d38
 
 
696f787
3ca1d38
9659593
3ca1d38
 
696f787
9659593
3ca1d38
 
696f787
9659593
3ca1d38
 
696f787
3ca1d38
9659593
3ca1d38
9659593
3ca1d38
 
 
696f787
9659593
3ca1d38
 
9659593
3ca1d38
 
 
696f787
9659593
3ca1d38
9659593
3ca1d38
 
9659593
3ca1d38
 
9659593
3ca1d38
 
696f787
3ca1d38
 
9659593
3ca1d38
 
 
9659593
3ca1d38
 
 
696f787
9659593
3ca1d38
 
 
 
 
 
 
 
696f787
3ca1d38
9659593
3ca1d38
 
696f787
9659593
3ca1d38
 
696f787
3ca1d38
 
9659593
3ca1d38
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
"""
MediGuard AI RAG-Helper - Evaluation System
5D Quality Assessment Framework

This module provides quality evaluation for RAG outputs using a 5-dimension framework:
1. Clinical Accuracy - Medical correctness (LLM-as-judge)
2. Evidence Grounding - Citation coverage (programmatic + LLM)
3. Actionability - Practical recommendations (LLM-as-judge)
4. Clarity - Communication quality (LLM-as-judge)
5. Safety Completeness - Safety alerts coverage (programmatic)

IMPORTANT LIMITATIONS:
- LLM-as-judge evaluations are non-deterministic (may vary between runs)
- Designed for offline batch evaluation, NOT production scoring
- Requires LLM API access (Groq or Gemini) for full evaluation
- Set EVALUATION_DETERMINISTIC=true for reproducible tests (uses heuristics)

Usage:
    from src.evaluation.evaluators import run_5d_evaluation

    result = run_5d_evaluation(final_response, pubmed_context)
    print(f"Average score: {result.average_score():.2f}")
"""

import json
import os
from typing import Any

from langchain_core.prompts import ChatPromptTemplate
from pydantic import BaseModel, Field

from src.llm_config import get_chat_model

# Set to True for deterministic evaluation (testing)
DETERMINISTIC_MODE = os.environ.get("EVALUATION_DETERMINISTIC", "false").lower() == "true"


class GradedScore(BaseModel):
    """Structured score with justification"""

    score: float = Field(description="Score from 0.0 to 1.0", ge=0.0, le=1.0)
    reasoning: str = Field(description="Justification for the score")


class EvaluationResult(BaseModel):
    """Complete 5D evaluation result"""

    clinical_accuracy: GradedScore
    evidence_grounding: GradedScore
    actionability: GradedScore
    clarity: GradedScore
    safety_completeness: GradedScore

    def to_vector(self) -> list[float]:
        """Extract scores as a vector for Pareto analysis"""
        return [
            self.clinical_accuracy.score,
            self.evidence_grounding.score,
            self.actionability.score,
            self.clarity.score,
            self.safety_completeness.score,
        ]

    def average_score(self) -> float:
        """Calculate average of all 5 dimensions"""
        scores = self.to_vector()
        return sum(scores) / len(scores) if scores else 0.0


# Evaluator 1: Clinical Accuracy (LLM-as-Judge)
def evaluate_clinical_accuracy(final_response: dict[str, Any], pubmed_context: str) -> GradedScore:
    """
    Evaluates if medical interpretations are accurate.
    Uses cloud LLM (Groq/Gemini) as expert judge.

    In DETERMINISTIC_MODE, uses heuristics instead.
    """
    # Deterministic mode for testing
    if DETERMINISTIC_MODE:
        return _deterministic_clinical_accuracy(final_response, pubmed_context)

    # Use cloud LLM for evaluation (FREE via Groq/Gemini)
    evaluator_llm = get_chat_model(temperature=0.0, json_mode=True)

    prompt = ChatPromptTemplate.from_messages(
        [
            (
                "system",
                """You are a medical expert evaluating clinical accuracy.
        
Evaluate the following clinical assessment:
- Are biomarker interpretations medically correct?
- Is the disease mechanism explanation accurate?
- Are the medical recommendations appropriate?

Score 1.0 = Perfectly accurate, no medical errors
Score 0.0 = Contains dangerous misinformation

Respond ONLY with valid JSON in this format:
{{"score": 0.85, "reasoning": "Your detailed justification here"}}
""",
            ),
            (
                "human",
                """Evaluate this clinical output:

**Patient Summary:**
{patient_summary}

**Prediction Explanation:**
{prediction_explanation}

**Clinical Recommendations:**
{recommendations}

**Scientific Context (Ground Truth):**
{context}
""",
            ),
        ]
    )

    chain = prompt | evaluator_llm
    result = chain.invoke(
        {
            "patient_summary": final_response["patient_summary"],
            "prediction_explanation": final_response["prediction_explanation"],
            "recommendations": final_response["clinical_recommendations"],
            "context": pubmed_context,
        }
    )

    # Parse JSON response
    try:
        content = result.content if isinstance(result.content, str) else str(result.content)
        parsed = json.loads(content)
        return GradedScore(score=parsed["score"], reasoning=parsed["reasoning"])
    except (json.JSONDecodeError, KeyError, TypeError):
        # Fallback if JSON parsing fails — use a conservative score to avoid inflating metrics
        return GradedScore(score=0.5, reasoning="Unable to parse LLM evaluation response; defaulting to neutral score.")


# Evaluator 2: Evidence Grounding (Programmatic + LLM)
def evaluate_evidence_grounding(final_response: dict[str, Any]) -> GradedScore:
    """
    Checks if all claims are backed by citations.
    Programmatic + LLM verification.
    """
    # Count citations
    pdf_refs = final_response["prediction_explanation"].get("pdf_references", [])
    citation_count = len(pdf_refs)

    # Check key drivers have evidence
    key_drivers = final_response["prediction_explanation"].get("key_drivers", [])
    drivers_with_evidence = sum(1 for d in key_drivers if d.get("evidence"))

    # Citation coverage score
    if len(key_drivers) > 0:
        coverage = drivers_with_evidence / len(key_drivers)
    else:
        coverage = 0.0

    # Base score from programmatic checks
    base_score = min(1.0, citation_count / 5.0) * 0.5 + coverage * 0.5

    reasoning = f"""
    Citations found: {citation_count}
    Key drivers with evidence: {drivers_with_evidence}/{len(key_drivers)}
    Citation coverage: {coverage:.1%}
    """

    return GradedScore(score=base_score, reasoning=reasoning.strip())


# Evaluator 3: Clinical Actionability (LLM-as-Judge)
def evaluate_actionability(final_response: dict[str, Any]) -> GradedScore:
    """
    Evaluates if recommendations are actionable and safe.
    Uses cloud LLM (Groq/Gemini) as expert judge.

    In DETERMINISTIC_MODE, uses heuristics instead.
    """
    # Deterministic mode for testing
    if DETERMINISTIC_MODE:
        return _deterministic_actionability(final_response)

    # Use cloud LLM for evaluation (FREE via Groq/Gemini)
    evaluator_llm = get_chat_model(temperature=0.0, json_mode=True)

    prompt = ChatPromptTemplate.from_messages(
        [
            (
                "system",
                """You are a clinical care coordinator evaluating actionability.

Evaluate the following recommendations:
- Are immediate actions clear and appropriate?
- Are lifestyle changes specific and practical?
- Are monitoring recommendations feasible?
- Are next steps clearly defined?

Score 1.0 = Perfectly actionable, clear next steps
Score 0.0 = Vague, impractical, or unsafe

Respond ONLY with valid JSON in this format:
{{"score": 0.90, "reasoning": "Your detailed justification here"}}
""",
            ),
            (
                "human",
                """Evaluate these recommendations:

**Immediate Actions:**
{immediate_actions}

**Lifestyle Changes:**
{lifestyle_changes}

**Monitoring:**
{monitoring}

**Confidence Assessment:**
{confidence}
""",
            ),
        ]
    )

    chain = prompt | evaluator_llm
    recs = final_response["clinical_recommendations"]
    result = chain.invoke(
        {
            "immediate_actions": recs.get("immediate_actions", []),
            "lifestyle_changes": recs.get("lifestyle_changes", []),
            "monitoring": recs.get("monitoring", []),
            "confidence": final_response["confidence_assessment"],
        }
    )

    # Parse JSON response
    try:
        parsed = json.loads(result.content if isinstance(result.content, str) else str(result.content))
        return GradedScore(score=parsed["score"], reasoning=parsed["reasoning"])
    except (json.JSONDecodeError, KeyError, TypeError):
        # Fallback if JSON parsing fails — use a conservative score to avoid inflating metrics
        return GradedScore(score=0.5, reasoning="Unable to parse LLM evaluation response; defaulting to neutral score.")


# Evaluator 4: Explainability Clarity (Programmatic)
def evaluate_clarity(final_response: dict[str, Any]) -> GradedScore:
    """
    Measures readability and patient-friendliness.
    Uses programmatic text analysis.

    In DETERMINISTIC_MODE, uses simple heuristics for reproducibility.
    """
    # Deterministic mode for testing
    if DETERMINISTIC_MODE:
        return _deterministic_clarity(final_response)

    try:
        import textstat

        has_textstat = True
    except ImportError:
        has_textstat = False

    # Get patient narrative
    narrative = final_response["patient_summary"].get("narrative", "")

    if has_textstat:
        # Calculate readability (Flesch Reading Ease)
        # Score 60-70 = Standard (8th-9th grade)
        # Score 50-60 = Fairly difficult (10th-12th grade)
        flesch_score = textstat.flesch_reading_ease(narrative)
        readability_score = min(1.0, flesch_score / 70.0)  # Normalize to 1.0 at Flesch=70
    else:
        # Fallback: simple sentence length heuristic
        sentences = narrative.split(".")
        avg_words = sum(len(s.split()) for s in sentences) / max(len(sentences), 1)
        # Optimal: 15-20 words per sentence
        if 15 <= avg_words <= 20:
            readability_score = 1.0
        elif avg_words < 15:
            readability_score = 0.9
        else:
            readability_score = max(0.5, 1.0 - (avg_words - 20) * 0.02)

    # Medical jargon detection (simple heuristic)
    medical_terms = [
        "pathophysiology",
        "etiology",
        "hemostasis",
        "coagulation",
        "thrombocytopenia",
        "erythropoiesis",
        "gluconeogenesis",
    ]
    jargon_count = sum(1 for term in medical_terms if term.lower() in narrative.lower())

    # Length check (too short = vague, too long = overwhelming)
    word_count = len(narrative.split())
    optimal_length = 50 <= word_count <= 150

    # Scoring
    jargon_penalty = max(0.0, 1.0 - (jargon_count * 0.2))
    length_score = 1.0 if optimal_length else 0.7

    final_score = readability_score * 0.5 + jargon_penalty * 0.3 + length_score * 0.2

    if has_textstat:
        reasoning = f"""
    Flesch Reading Ease: {flesch_score:.1f} (Target: 60-70)
    Medical jargon terms: {jargon_count}
    Word count: {word_count} (Optimal: 50-150)
    Readability subscore: {readability_score:.2f}
    """
    else:
        reasoning = f"""
    Readability (heuristic): {readability_score:.2f}
    Medical jargon terms: {jargon_count}
    Word count: {word_count} (Optimal: 50-150)
    Note: textstat not available, using fallback metrics
    """

    return GradedScore(score=final_score, reasoning=reasoning.strip())


# Evaluator 5: Safety & Completeness (Programmatic)
def evaluate_safety_completeness(final_response: dict[str, Any], biomarkers: dict[str, float]) -> GradedScore:
    """
    Checks if all safety concerns are flagged.
    Programmatic validation.
    """
    from src.biomarker_validator import BiomarkerValidator

    # Initialize validator
    validator = BiomarkerValidator()

    # Count out-of-range biomarkers
    out_of_range_count = 0
    critical_count = 0

    for name, value in biomarkers.items():
        result = validator.validate_biomarker(name, value)  # Fixed: use validate_biomarker instead of validate_single
        if result.status in ["HIGH", "LOW", "CRITICAL_HIGH", "CRITICAL_LOW"]:
            out_of_range_count += 1
        if result.status in ["CRITICAL_HIGH", "CRITICAL_LOW"]:
            critical_count += 1

    # Count safety alerts in output
    safety_alerts = final_response.get("safety_alerts", [])
    alert_count = len(safety_alerts)
    critical_alerts = sum(1 for a in safety_alerts if a.get("severity") == "CRITICAL")

    # Check if all critical values have alerts
    critical_coverage = critical_alerts / critical_count if critical_count > 0 else 1.0

    # Check for disclaimer
    has_disclaimer = "disclaimer" in final_response.get("metadata", {})

    # Check for uncertainty acknowledgment
    limitations = final_response["confidence_assessment"].get("limitations", [])
    acknowledges_uncertainty = len(limitations) > 0

    # Scoring
    alert_score = min(1.0, alert_count / max(1, out_of_range_count))
    critical_score = min(1.0, critical_coverage)
    disclaimer_score = 1.0 if has_disclaimer else 0.0
    uncertainty_score = 1.0 if acknowledges_uncertainty else 0.5

    final_score = min(
        1.0, (alert_score * 0.4 + critical_score * 0.3 + disclaimer_score * 0.2 + uncertainty_score * 0.1)
    )

    reasoning = f"""
    Out-of-range biomarkers: {out_of_range_count}
    Critical values: {critical_count}
    Safety alerts generated: {alert_count}
    Critical alerts: {critical_alerts}
    Critical coverage: {critical_coverage:.1%}
    Has disclaimer: {has_disclaimer}
    Acknowledges uncertainty: {acknowledges_uncertainty}
    """

    return GradedScore(score=final_score, reasoning=reasoning.strip())


# Master Evaluation Function
def run_full_evaluation(
    final_response: dict[str, Any], agent_outputs: list[Any], biomarkers: dict[str, float]
) -> EvaluationResult:
    """
    Orchestrates all 5 evaluators and returns complete assessment.
    """
    print("=" * 70)
    print("RUNNING 5D EVALUATION GAUNTLET")
    print("=" * 70)

    # Extract context from agent outputs
    pubmed_context = ""
    for output in agent_outputs:
        if output.agent_name == "Disease Explainer":
            findings = output.findings
            if isinstance(findings, dict):
                pubmed_context = findings.get("mechanism_summary", "") or findings.get("pathophysiology", "")
            elif isinstance(findings, str):
                pubmed_context = findings
            else:
                pubmed_context = str(findings)
            break

    # Run all evaluators
    print("\n1. Evaluating Clinical Accuracy...")
    clinical_accuracy = evaluate_clinical_accuracy(final_response, pubmed_context)

    print("2. Evaluating Evidence Grounding...")
    evidence_grounding = evaluate_evidence_grounding(final_response)

    print("3. Evaluating Clinical Actionability...")
    actionability = evaluate_actionability(final_response)

    print("4. Evaluating Explainability Clarity...")
    clarity = evaluate_clarity(final_response)

    print("5. Evaluating Safety & Completeness...")
    safety_completeness = evaluate_safety_completeness(final_response, biomarkers)

    print("\n" + "=" * 70)
    print("EVALUATION COMPLETE")
    print("=" * 70)

    return EvaluationResult(
        clinical_accuracy=clinical_accuracy,
        evidence_grounding=evidence_grounding,
        actionability=actionability,
        clarity=clarity,
        safety_completeness=safety_completeness,
    )


# ---------------------------------------------------------------------------
# Deterministic Evaluation Functions (for testing)
# ---------------------------------------------------------------------------


def _deterministic_clinical_accuracy(final_response: dict[str, Any], pubmed_context: str) -> GradedScore:
    """Heuristic-based clinical accuracy (deterministic)."""
    score = 0.5
    reasons = []

    # Check if response has expected structure
    if final_response.get("patient_summary"):
        score += 0.1
        reasons.append("Has patient summary")

    if final_response.get("prediction_explanation"):
        score += 0.1
        reasons.append("Has prediction explanation")

    if final_response.get("clinical_recommendations"):
        score += 0.1
        reasons.append("Has clinical recommendations")

    # Check for citations
    pred = final_response.get("prediction_explanation", {})
    if isinstance(pred, dict):
        refs = pred.get("pdf_references", [])
        if refs:
            score += min(0.2, len(refs) * 0.05)
            reasons.append(f"Has {len(refs)} citations")

    return GradedScore(score=min(1.0, score), reasoning="[DETERMINISTIC] " + "; ".join(reasons))


def _deterministic_actionability(final_response: dict[str, Any]) -> GradedScore:
    """Heuristic-based actionability (deterministic)."""
    score = 0.5
    reasons = []

    recs = final_response.get("clinical_recommendations", {})
    if isinstance(recs, dict):
        if recs.get("immediate_actions"):
            score += 0.15
            reasons.append("Has immediate actions")
        if recs.get("lifestyle_changes"):
            score += 0.15
            reasons.append("Has lifestyle changes")
        if recs.get("monitoring"):
            score += 0.1
            reasons.append("Has monitoring recommendations")

    return GradedScore(
        score=min(1.0, score),
        reasoning="[DETERMINISTIC] " + "; ".join(reasons) if reasons else "[DETERMINISTIC] Missing recommendations",
    )


def _deterministic_clarity(final_response: dict[str, Any]) -> GradedScore:
    """Heuristic-based clarity (deterministic)."""
    score = 0.5
    reasons = []

    summary = final_response.get("patient_summary", "")
    if isinstance(summary, str):
        word_count = len(summary.split())
        if 50 <= word_count <= 300:
            score += 0.2
            reasons.append(f"Summary length OK ({word_count} words)")
        elif word_count > 0:
            score += 0.1
            reasons.append("Has summary")

    # Check for structured output
    if final_response.get("biomarker_flags"):
        score += 0.15
        reasons.append("Has biomarker flags")

    if final_response.get("key_findings"):
        score += 0.15
        reasons.append("Has key findings")

    return GradedScore(
        score=min(1.0, score),
        reasoning="[DETERMINISTIC] " + "; ".join(reasons) if reasons else "[DETERMINISTIC] Limited structure",
    )