File size: 9,262 Bytes
404c45f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
158878f
 
 
 
 
 
 
404c45f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
80c426b
404c45f
 
 
 
158878f
404c45f
 
 
 
 
 
 
 
80c426b
404c45f
 
 
 
158878f
404c45f
 
 
 
 
 
 
 
80c426b
404c45f
 
 
 
158878f
404c45f
 
 
158878f
404c45f
 
 
 
 
 
 
 
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
"""
Heuristic Baseline
===================
A deterministic rule-based baseline agent for all 3 tasks.
Used by the /baseline endpoint for fast, reproducible scoring.
The LLM-based baseline (GroqCloud API) is in scripts/baseline_inference.py.
"""
from __future__ import annotations

from typing import Any, Dict

from models import (
    AdverseEventTriageAction,
    CausalityAssessment,
    DeviationType,
    ProtocolDeviationAction,
    SafetyNarrativeAction,
    TaskID,
    TriageAction,
)
from server.environment import ClinicalTrialEnvironment
from tasks.case_bank import AE_CASES, DEVIATION_CASES, NARRATIVE_CASES


_SCORE_EPS = 1e-3


def _clamp_open_score(value: float) -> float:
    return max(_SCORE_EPS, min(1.0 - _SCORE_EPS, float(value)))


# ─────────────────────────────────────────
# HEURISTIC AGENTS (rule-based)
# ─────────────────────────────────────────

def _heuristic_ae_triage(case: Dict[str, Any]) -> TriageAction:
    """
    Simple keyword/lab-value heuristic for AE triage.
    Gets ~70% on easy cases, ~45% on hard cases.
    """
    narrative = case["narrative"].lower()
    ae_desc = case["ae_description"].lower()

    # Severity heuristic
    if any(kw in narrative for kw in ["fatal", "death", "died"]):
        severity = "fatal"
        timeline = "7-day"
        is_serious = True
    elif any(kw in narrative for kw in ["stemi", "cardiac arrest", "icu", "intensive care", "life-threatening"]):
        severity = "life_threatening"
        timeline = "7-day"
        is_serious = True
    elif any(kw in narrative for kw in ["hospitali", "encephalopathy", "grade 3", "severe"]):
        severity = "severe"
        timeline = "15-day"
        is_serious = True
    elif any(kw in narrative for kw in ["grade 2", "moderate", "nausea"]):
        severity = "moderate"
        timeline = "routine"
        is_serious = False
    else:
        severity = "mild"
        timeline = "routine"
        is_serious = False

    # MedDRA SOC heuristic
    if any(kw in ae_desc for kw in ["cardiac", "myocardial", "stemi", "heart"]):
        soc = "Cardiac disorders"
        pt = "Myocardial infarction"
    elif any(kw in ae_desc for kw in ["nausea", "vomiting", "gastrointestinal"]):
        soc = "Gastrointestinal disorders"
        pt = "Nausea"
    elif any(kw in ae_desc for kw in ["encephalopathy", "neurological", "nervous"]):
        soc = "Nervous system disorders"
        pt = "Encephalopathy"
    else:
        soc = "General disorders"
        pt = "Adverse event"

    return TriageAction(
        task_id=TaskID.ADVERSE_EVENT_TRIAGE,
        ae_triage=AdverseEventTriageAction(
            severity_classification=severity,
            reporting_timeline=timeline,
            meddra_soc=soc,
            meddra_preferred_term=pt,
            is_serious=is_serious,
            rationale="Heuristic baseline classification based on keyword matching.",
        ),
    )


def _heuristic_deviation_audit(case: Dict[str, Any]) -> TriageAction:
    """Heuristic deviation audit based on finding severity keywords."""
    findings = case["findings"]
    high_risk_keywords = [
        "eligibility", "blinding", "unblind", "sae report", "integrity",
        "data", "enroll", "hospitali", "ip accountability", "unaccounted",
        "endpoint", "consent", "delegate"
    ]

    risk_count = 0
    flagged_ids = []
    for f in findings:
        desc_lower = f["description"].lower()
        cat_lower = f["category"].lower()
        if any(kw in desc_lower or kw in cat_lower for kw in high_risk_keywords):
            risk_count += 1
            flagged_ids.append(f["id"])

    is_major = risk_count >= 2
    capa = risk_count >= 2
    risk_score = min(10.0, risk_count * 2.5 + case.get("prior_deviations", 0) * 0.3)

    return TriageAction(
        task_id=TaskID.PROTOCOL_DEVIATION_AUDIT,
        deviation_audit=ProtocolDeviationAction(
            deviation_type=DeviationType.MAJOR if is_major else DeviationType.MINOR,
            capa_required=capa,
            site_risk_score=risk_score,
            flagged_finding_ids=flagged_ids,
            recommended_action="Immediate escalation to CRA and Sponsor QA team for review." if is_major else "Document and include in next monitoring report.",
        ),
    )


def _heuristic_narrative(case: Dict[str, Any]) -> TriageAction:
    """Heuristic narrative generation β€” structured template filling."""
    dem = case["patient_demographics"]
    ae = case["adverse_event"]
    labs = case["lab_values_timeline"]
    conmeds = case["concomitant_medications"]

    conmed_str = "; ".join(f"{m['name']} {m['dose']}" for m in conmeds)
    lab_str = "; ".join(
        f"{l['date']}: INR {l.get('INR', 'N/A')}, Hgb {l.get('Hgb_g_dL', 'N/A')} g/dL"
        for l in labs
    )

    narrative = (
        f"A {dem['age']}-year-old {dem['sex']} patient enrolled in a clinical study "
        f"received {case['study_drug']}. "
        f"Relevant medical history includes: {'; '.join(case['medical_history'])}. "
        f"Concomitant medications: {conmed_str}. "
        f"On {ae['onset_date']}, the patient developed {ae['term']} (MedDRA: {ae['meddra_soc']} / {ae['meddra_pt']}), "
        f"meeting seriousness criteria of {', '.join(ae['seriousness_criteria'])}. "
        f"Laboratory values over time: {lab_str}. "
        f"Notable INR elevation to {labs[-2].get('INR', 'N/A') if len(labs) >= 2 else 'N/A'} was observed prior to the event, "
        f"suggesting a potential drug-drug interaction between {case['study_drug']} and warfarin. "
        f"Action taken: {case['action_taken']}. "
        f"Dechallenge was positive β€” the event resolved following drug discontinuation. "
        f"Outcome at last follow-up: {case['outcome_at_last_followup']}. "
        f"Causality assessment: The event is considered probably related to the study drug, "
        f"given the temporal relationship, positive dechallenge, and the plausible pharmacokinetic "
        f"interaction with warfarin resulting in supratherapeutic INR levels."
    )

    return TriageAction(
        task_id=TaskID.SAFETY_NARRATIVE_GENERATION,
        safety_narrative=SafetyNarrativeAction(
            narrative_text=narrative,
            causality_assessment=CausalityAssessment.PROBABLY_RELATED,
            key_temporal_flags=[
                "INR elevation 2 days prior to event",
                "onset after dose initiation day 14",
                "positive dechallenge on drug discontinuation",
            ],
            dechallenge_positive=True,
            rechallenge_positive=None,
        ),
    )


# ─────────────────────────────────────────
# MAIN RUNNER
# ─────────────────────────────────────────

def run_heuristic_baseline() -> Dict[str, Any]:
    """Run heuristic baseline on all 3 tasks and return scores."""
    env = ClinicalTrialEnvironment()
    results: Dict[str, Any] = {
        "baseline_type": "heuristic",
        "description": "Rule-based keyword matching baseline β€” establishes lower bound.",
        "tasks": {},
    }

    # Task 1: AE Triage
    env.reset(task_id=TaskID.ADVERSE_EVENT_TRIAGE)
    ae_rewards = []
    for case in AE_CASES:
        action = _heuristic_ae_triage(case)
        result = env.step(action)
        ae_rewards.append(_clamp_open_score(float(result.reward)))
        if result.done:
            break

    results["tasks"][TaskID.ADVERSE_EVENT_TRIAGE] = {
        "mean_reward": round(_clamp_open_score(sum(ae_rewards) / len(ae_rewards)), 4) if ae_rewards else _clamp_open_score(_SCORE_EPS),
    }

    # Task 2: Protocol Deviation Audit
    env.reset(task_id=TaskID.PROTOCOL_DEVIATION_AUDIT)
    dev_rewards = []
    for case in DEVIATION_CASES:
        action = _heuristic_deviation_audit(case)
        result = env.step(action)
        dev_rewards.append(_clamp_open_score(float(result.reward)))
        if result.done:
            break

    results["tasks"][TaskID.PROTOCOL_DEVIATION_AUDIT] = {
        "mean_reward": round(_clamp_open_score(sum(dev_rewards) / len(dev_rewards)), 4) if dev_rewards else _clamp_open_score(_SCORE_EPS),
    }

    # Task 3: Safety Narrative
    env.reset(task_id=TaskID.SAFETY_NARRATIVE_GENERATION)
    nr_rewards = []
    for case in NARRATIVE_CASES:
        action = _heuristic_narrative(case)
        result = env.step(action)
        nr_rewards.append(_clamp_open_score(float(result.reward)))
        if result.done:
            break

    results["tasks"][TaskID.SAFETY_NARRATIVE_GENERATION] = {
        "mean_reward": round(_clamp_open_score(sum(nr_rewards) / len(nr_rewards)), 4) if nr_rewards else _clamp_open_score(_SCORE_EPS),
    }

    all_means = [v["mean_reward"] for v in results["tasks"].values()]
    results["overall_mean_reward"] = round(_clamp_open_score(sum(all_means) / len(all_means)), 4)

    return results


if __name__ == "__main__":
    import json
    results = run_heuristic_baseline()
    print(json.dumps(results, indent=2))