generate_dataset.py

#!/usr/bin/env python3
"""
AI & Clinical Decision Support Dataset
==========================================
Each record = ONE AI/CDSS deployment or interaction event.

Literature-Grounded Parameterization:
[1] Wahl B et al. (2018). AI for global health in SSA.
    BMJ Glob Health. DOI: 10.1136/bmjgh-2018-000798
    - AI applications: diagnosis, triage, drug discovery
    - Most SSA AI health projects: pilot stage
    - Chest X-ray AI most deployed (TB screening)

[2] Topol EJ (2019). High-performance medicine with AI.
    Nat Med. DOI: 10.1038/s41591-018-0300-7
    - CDSS improves diagnostic accuracy 10-30%
    - AI radiology, pathology, dermatology leading domains
    - Validation in local populations critical

[3] Mollura DJ et al. (2020). AI for radiology in LMICs.
    Radiology. DOI: 10.1148/radiol.2020201434
    - AI TB detection: sensitivity 90-98%
    - Diabetic retinopathy screening AI deployed
    - Barriers: data quality, infrastructure, regulation
"""
import numpy as np, pandas as pd, argparse, os

SCENARIOS = {
    'ai_deployed': {
        'exemplar': 'South Africa/Kenya/Rwanda (urban)',
        'ai_available': 0.12,
        'cdss_available': 0.15,
        'ai_radiology': 0.08,
        'ai_pathology': 0.03,
        'ai_dermatology': 0.02,
        'ai_triage': 0.05,
        'recommendation_followed': 0.55,
        'diagnostic_accuracy_boost': 0.20,
        'user_trust': 0.50,
        'internet_required': 0.70,
        'validated_local': 0.15,
    },
    'ai_pilot': {
        'exemplar': 'Kenya (district)/Ghana/Tanzania/Ethiopia',
        'ai_available': 0.03,
        'cdss_available': 0.05,
        'ai_radiology': 0.02,
        'ai_pathology': 0.005,
        'ai_dermatology': 0.005,
        'ai_triage': 0.01,
        'recommendation_followed': 0.35,
        'diagnostic_accuracy_boost': 0.12,
        'user_trust': 0.30,
        'internet_required': 0.80,
        'validated_local': 0.05,
    },
    'ai_absent': {
        'exemplar': 'DRC/CAR/Sierra Leone/Niger (rural)',
        'ai_available': 0.005,
        'cdss_available': 0.01,
        'ai_radiology': 0.002,
        'ai_pathology': 0.001,
        'ai_dermatology': 0.001,
        'ai_triage': 0.002,
        'recommendation_followed': 0.20,
        'diagnostic_accuracy_boost': 0.05,
        'user_trust': 0.15,
        'internet_required': 0.90,
        'validated_local': 0.01,
    },
}

def generate_dataset(n=10000, seed=42, scenario='ai_pilot'):
    rng = np.random.default_rng(seed)
    sc = SCENARIOS[scenario]
    records = []
    for idx in range(n):
        rec = {'id': idx + 1}
        rec['facility_level'] = rng.choice(['tertiary','regional','district','health_centre','clinic','community'],
            p=[0.08, 0.12, 0.25, 0.30, 0.15, 0.10])
        rec['location'] = rng.choice(['urban','peri_urban','rural','remote'], p=[0.25, 0.20, 0.30, 0.25])

        tier = 1.5 if rec['facility_level'] in ['tertiary','regional'] else 0.6
        urban = 1.3 if rec['location'] in ['urban','peri_urban'] else 0.7

        # Infrastructure
        rec['internet_available'] = 1 if rng.random() < min(0.90, 0.35*urban*tier) else 0
        rec['power_reliable'] = 1 if rng.random() < min(0.90, 0.40*urban) else 0
        rec['smartphone_available'] = 1 if rng.random() < min(0.90, 0.50*urban) else 0

        # AI/CDSS availability
        rec['ai_tool_available'] = 1 if rng.random() < min(0.80, sc['ai_available']*tier*urban) else 0
        rec['cdss_available'] = 1 if rng.random() < min(0.80, sc['cdss_available']*tier*urban) else 0

        # AI domain
        rec['ai_domain'] = rng.choice([
            'tb_xray_screening','diabetic_retinopathy','dermatology','pathology',
            'triage_severity','drug_interaction','maternal_risk','malaria_microscopy',
            'ecg_interpretation','clinical_guidelines','other'],
            p=[0.18, 0.10, 0.08, 0.06, 0.12, 0.08, 0.10, 0.08, 0.05, 0.10, 0.05])

        # AI tool specifics
        rec['ai_type'] = rng.choice(['image_recognition','nlp','predictive_model','rule_based','hybrid'],
            p=[0.30, 0.10, 0.20, 0.25, 0.15])
        dp = np.array([0.30*urban, 0.20, max(0.05, 0.30-0.15*urban), 0.15])
        dp = dp / dp.sum()
        rec['deployment_mode'] = rng.choice(['cloud','edge_device','offline','hybrid'], p=dp)

        rec['open_source'] = 1 if rng.random() < 0.40 else 0
        rec['regulatory_approved'] = 1 if (rec['ai_tool_available'] and rng.random() < 0.15) else 0
        rec['validated_local_population'] = 1 if rng.random() < sc['validated_local'] else 0

        # Usage
        rec['ai_used_this_encounter'] = 1 if (rec['ai_tool_available'] and rng.random() < 0.60) else 0
        rec['recommendation_generated'] = 1 if (rec['ai_used_this_encounter'] and rng.random() < 0.80) else 0
        rec['recommendation_followed'] = 1 if (rec['recommendation_generated'] and rng.random() < sc['recommendation_followed']) else 0
        rec['recommendation_overridden'] = 1 if (rec['recommendation_generated'] and not rec['recommendation_followed'] and rng.random() < 0.50) else 0
        rec['override_reason'] = rng.choice(['clinical_judgment','patient_preference','data_quality','not_applicable','other'],
            p=[0.40, 0.15, 0.20, 0.15, 0.10]) if rec['recommendation_overridden'] else 'na'

        # Outcomes
        rec['diagnostic_accuracy_improved'] = 1 if (rec['ai_used_this_encounter'] and rng.random() < sc['diagnostic_accuracy_boost']) else 0
        rec['time_saved_minutes'] = max(0, int(rng.normal(10, 5))) if rec['ai_used_this_encounter'] else 0
        rec['patient_outcome_improved'] = 1 if (rec['recommendation_followed'] and rng.random() < 0.30) else 0
        rec['false_positive'] = 1 if (rec['ai_used_this_encounter'] and rng.random() < 0.08) else 0
        rec['false_negative'] = 1 if (rec['ai_used_this_encounter'] and rng.random() < 0.05) else 0

        # User experience
        rec['provider_trust_ai'] = int(np.clip(rng.normal(6 if sc['user_trust']>0.4 else 4, 2), 1, 10))
        rec['training_received'] = 1 if rng.random() < sc['ai_available']*3 else 0
        rec['ease_of_use_score'] = int(np.clip(rng.normal(6 if sc['ai_available']>0.05 else 4, 2), 1, 10))

        # Barriers
        rec['barrier_infrastructure'] = 1 if not rec['internet_available'] else 0
        rec['barrier_trust'] = 1 if rng.random() < max(0.10, 0.50-sc['user_trust']) else 0
        rec['barrier_training'] = 1 if not rec['training_received'] else 0
        rec['barrier_regulation'] = 1 if rng.random() < 0.30 else 0
        rec['barrier_data_quality'] = 1 if rng.random() < 0.35 else 0
        rec['barrier_cost'] = 1 if rng.random() < 0.40 else 0
        rec['barrier_bias_concern'] = 1 if rng.random() < 0.20 else 0

        rec['year'] = rng.choice([2019,2020,2021,2022,2023], p=[0.05,0.12,0.20,0.28,0.35])
        records.append(rec)

    df = pd.DataFrame(records)
    print(f"\n{'='*60}\nAI & Clinical Decision Support — {scenario} ({sc['exemplar']})")
    print(f"  AI available: {df['ai_tool_available'].mean()*100:.1f}% | Used: {df['ai_used_this_encounter'].mean()*100:.1f}%")
    print(f"  Rec followed: {df['recommendation_followed'].mean()*100:.1f}% | Accuracy boost: {df['diagnostic_accuracy_improved'].mean()*100:.1f}%")
    print(f"  Trust score: {df['provider_trust_ai'].mean():.1f}/10 | Validated local: {df['validated_local_population'].mean()*100:.1f}%")
    return df

if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--all-scenarios', action='store_true')
    parser.add_argument('--n', type=int, default=10000)
    parser.add_argument('--seed', type=int, default=42)
    args = parser.parse_args()
    os.makedirs('data', exist_ok=True)
    if args.all_scenarios:
        for sc in SCENARIOS:
            df = generate_dataset(n=args.n, seed=args.seed, scenario=sc)
            df.to_csv(os.path.join('data', f'ai_cdss_{sc}.csv'), index=False)
            print(f"  -> Saved\n")
    else:
        df = generate_dataset(n=args.n, seed=args.seed)
        df.to_csv(os.path.join('data', 'ai_cdss_ai_pilot.csv'), index=False)