"""Triangulation-targeted 후속 가설 + 8가지 design 카드. discordance/missing-design 갭에 따라 다음 design 8종 중 최적 추천: 1. Active comparator new-user (target trial emulation) 2. Multivariable MR (BMI + WHR + body fat %) 3. Negative control outcome / negative control exposure 4. 2×2 factorial (예: GLP-1RA × exercise) 5. Within-subject crossover 6. Bariatric preference-based IV 7. DIO animal weight-matched control 8. Mediation N-of-1 / E-N-of-1 """ from __future__ import annotations from typing import Dict, List def design_card(design_type: str, intervention: str, outcome: str, rationale: str) -> Dict: """후속 design 카드 1개 생성.""" templates = { "active_comparator_new_user": { "primary_endpoint": "3P-MACE / outcome 특이", "sample_size_estimate": "n=10,000~50,000 (claims/EHR)", "follow_up_years": "2~5", "key_hypothesis": ( f"{intervention} vs DPP-4i (active comparator) new-user " f"propensity-matched analysis로 confounding by indication 최소화." ), "multivariable_MR_instruments": "N/A (real-world emulation)", "key_threats": "residual unmeasured confounding (체중·HbA1c)", "minimum_data": "처방 시작일·index date·comorbidity·BMI", }, "multivariable_MR_BMI_WHR_BF": { "primary_endpoint": "outcome OR per SD adiposity phenotype", "sample_size_estimate": "n=500,000 (UK Biobank, FinnGen, KoGES)", "follow_up_years": "lifetime exposure", "key_hypothesis": ( f"BMI, WHR, body fat % 동시 IV로 {intervention}이 영향 주는 " f"adiposity 표현형을 conditional 분리하여 {outcome} 효과의 " f"adipose-distribution-specific contribution 추정." ), "multivariable_MR_instruments": ( "GIANT BMI 941 SNP, UK Biobank WHR 346 SNP, body fat % 425 SNP. " "MVMR-Egger / MVMR-IVW / MR-PRESSO sensitivity." ), "key_threats": "weak instrument bias (conditional F<10), horizontal pleiotropy", "minimum_data": "GWAS summary stats + outcome GWAS", }, "negative_control_outcome": { "primary_endpoint": "결과와 무관한 outcome (예: 사고사)", "sample_size_estimate": "동일 cohort 활용", "follow_up_years": "동일", "key_hypothesis": ( f"{intervention} 처방자의 사고사·교통사고 등 " "biology-unrelated outcome에 'effect'가 있으면 healthy-user bias 시사." ), "multivariable_MR_instruments": "N/A", "key_threats": "negative control 선택의 misspecification", "minimum_data": "outcome 분류 ICD", }, "factorial_2x2": { "primary_endpoint": "체성분·outcome composite", "sample_size_estimate": "n=400~1,200", "follow_up_years": "1~2", "key_hypothesis": ( f"{intervention} × resistance exercise 2×2 factorial로 " "sarcopenia/lean mass loss를 운동이 부분 상쇄하는지 검증. " "interaction term 통계적 유의성." ), "multivariable_MR_instruments": "N/A", "key_threats": "adherence (특히 운동 군)", "minimum_data": "DXA, 6MWT, grip strength", }, "within_subject_crossover": { "primary_endpoint": "단기 surrogate (HbA1c, BP, 체중)", "sample_size_estimate": "n=60~200", "follow_up_years": "0.5~1", "key_hypothesis": ( f"on-{intervention}/off-{intervention} 4기 crossover로 " "individual-level treatment effect 분포 추정." ), "multivariable_MR_instruments": "N/A", "key_threats": "carryover (GLP-1 long t1/2), period bias", "minimum_data": "high-frequency biomarker", }, "bariatric_preference_IV": { "primary_endpoint": outcome, "sample_size_estimate": "n>5,000 multi-center", "follow_up_years": "5~20", "key_hypothesis": ( "center-level bariatric vs medical 선호도를 IV로 활용, " f"{intervention} (bariatric)의 conditional effect를 selection bias 보정 후 추정." ), "multivariable_MR_instruments": "preference-IV first-stage F>10 필요", "key_threats": "IV exclusion 위반 (center가 다른 영향)", "minimum_data": "multi-center cohort, center ID, outcome", }, "DIO_animal_weight_matched": { "primary_endpoint": "tissue-specific (간·근·뇌·심)", "sample_size_estimate": "n=80~150 mice", "follow_up_years": "0.2~0.6", "key_hypothesis": ( f"DIO mouse에 {intervention} vs vehicle + pair-fed weight-matched control로 " "weight-independent direct effect (GLP-1R 신경/심장) 분리." ), "multivariable_MR_instruments": "N/A", "key_threats": "species extrapolation, pair-feeding stress", "minimum_data": "tissue histology, RNA-seq", }, "mediation_N_of_1": { "primary_endpoint": "individual mediated/direct decomposition", "sample_size_estimate": "n=1 × multiple periods", "follow_up_years": "0.5~1", "key_hypothesis": ( f"단일 환자에서 {intervention} dose adjustment으로 " "weight loss와 outcome 변화의 individual mediation 추정." ), "multivariable_MR_instruments": "N/A", "key_threats": "single subject 일반화 불가", "minimum_data": "CGM/wearable, weekly biomarker", }, } t = templates.get(design_type, {}) return { "type": design_type, "intervention": intervention, "outcome": outcome, "rationale": rationale, **t, } def recommend_designs(intervention: str, outcome: str) -> List[Dict]: """현재 데이터의 갭에 따라 8 design 중 우선순위 추천.""" from .grid import get_pair_rows rows = get_pair_rows(intervention, outcome) have = set(r["design"] for r in rows) recommendations = [] # RCT 없으면 active comparator if "RCT" not in have: recommendations.append(design_card( "active_comparator_new_user", intervention, outcome, "RCT가 없거나 비실현적인 outcome → target trial emulation 우선." )) # MVMR 없으면 추천 if "multivariable_MR" not in have: recommendations.append(design_card( "multivariable_MR_BMI_WHR_BF", intervention, outcome, "BMI MR만 있고 WHR/BF% 분리 없음 → adiposity 표현형 contribution 추정 필요." )) # observational discordant이거나 missing if "observational" in have: recommendations.append(design_card( "negative_control_outcome", intervention, outcome, "Observational 효과가 RCT보다 크면 healthy-user bias 검증 필요." )) # bariatric 큰 효과 if any("bariatric" in r["intervention"] for r in rows) and "bariatric_natural" in have: recommendations.append(design_card( "bariatric_preference_IV", intervention, outcome, "Bariatric vs medical comparison에서 selection bias 보정." )) # sarcopenia/lean mass outcome if outcome.lower() in ("sarcopenia", "hip_fracture", "osteoarthritis"): recommendations.append(design_card( "factorial_2x2", intervention, outcome, "Lean mass loss를 운동이 mitigate하는지 factorial로 검증." )) # DIO animal 없으면 if "DIO_animal" not in have: recommendations.append(design_card( "DIO_animal_weight_matched", intervention, outcome, "Weight-independent direct effect 분리 위해 pair-fed control 필요." )) # within-subject 없으면 if "within_subject_crossover" not in have: recommendations.append(design_card( "within_subject_crossover", intervention, outcome, "Individual-level treatment effect heterogeneity 평가." )) # mediation N-of-1 recommendations.append(design_card( "mediation_N_of_1", intervention, outcome, "Personalized mediation: dose adjustment → mediated/direct fraction." )) return recommendations[:8] if __name__ == "__main__": cards = recommend_designs("semaglutide", "CV death") for c in cards: print(c["type"], "—", c["rationale"][:80])