GPT Integration in Miami, FL

Healthcare / Multi-Facility System

A South Florida health system (11 facilities, 1,800 physicians, 520,000 active patients — 58% preferring Spanish-language communication) was struggling with patient communication quality and consistency. Communication gaps: after-visit summaries: generated through the EMR's (Epic) built-in tools — technically accurate but written in clinical language that patients struggled to understand. Reading level: averaging grade 12 (college level), significantly above the recommended grade 6-8 for patient materials. Spanish summaries: produced through the EMR's translation module — grammatically correct but lacking the natural flow and cultural sensitivity of native Spanish communication. Patient comprehension: surveys showing only 42% of patients fully understanding their after-visit instructions. Care instruction adherence: directly impacted by comprehension — patients who didn't understand their care instructions showing 34% lower adherence rates. Patient education: generic materials selected from a library of pre-written content — not personalised to the patient's specific condition, medications, or cultural context. The materials: available in English and Spanish but not adapted for the dietary traditions, health beliefs, and lifestyle patterns of Miami's diverse patient communities. Outreach limitations: preventive care outreach using template-based messaging — the same cancer screening reminder sent to every eligible patient regardless of language, cultural context, or communication preference. Outreach response rates: 12% for English-language messages, 8% for Spanish-language messages — the generic messaging failing to motivate action. Documentation burden: physicians spending an average of 2.1 hours daily on clinical documentation — notes, referral letters, prior authorisation justifications, and patient correspondence. The burden: cited by 67% of physicians as the primary source of professional dissatisfaction. Technology limitation: the EMR offered limited AI features — basic auto-text templates and structured note entry, but no generative AI capability for personalised patient communication or documentation assistance.

We integrated GPT-4 into the health system's Epic-based workflow over 14 weeks. After-visit summary enhancement (weeks 1-6). GPT integration: connected to Epic through the system's FHIR APIs — pulling visit data (diagnosis, treatment plan, medications, follow-up requirements, and warning signs) and generating personalised after-visit summaries. Readability: GPT prompted to generate summaries at a 6th-grade reading level — explaining medical concepts in plain language with analogies and examples relevant to the patient's context. Bilingual generation: summaries generated natively in the patient's preferred language — not translated English but natural Spanish communication using culturally appropriate medical terminology. A patient seen for diabetes management: receiving a summary that explains blood sugar targets using familiar food references (Cuban rice and beans, plantains, café con leche) rather than abstract glycemic index numbers. Cultural adaptation: the GPT recognising patient demographic data and adapting content accordingly — a Haitian Creole-speaking patient receiving simplified English with cultural health context; a Brazilian patient receiving Portuguese content where available or adapted Spanish. Clinical validation: every generated summary validated against structured clinical data — ensuring no hallucinated medications, incorrect dosages, or fabricated instructions. Validation: automated comparison between GPT output and source EMR data flagging any discrepancies for physician review. Personalised patient education (weeks 4-10). GPT-generated education: personalised health education materials generated from the patient's clinical profile — condition-specific information, medication guides, dietary recommendations, and lifestyle modifications. Dietary adaptation: nutrition guidance adapted to the patient's cultural dietary patterns — not generic "eat more vegetables" but specific guidance like "when preparing arroz con pollo, consider using brown rice and adding extra peppers for fibre" for a patient managing diabetes within a Cuban-American dietary tradition. Medication guides: plain-language medication information including purpose, dosage, timing, common side effects, and interactions — generated for each patient's specific medication combination rather than generic drug information sheets. Chronic disease management: ongoing education series for patients with chronic conditions — weekly or monthly content covering different aspects of disease management, personalised to the patient's progress, lab results, and stated challenges. Care gap outreach enhancement (weeks 7-12). Personalised outreach: GPT generating individualised outreach messages for preventive care — cancer screening, vaccination, wellness visits, and chronic disease follow-up. The messages: personalised based on patient demographics, health history, and prior interaction patterns. Cultural framing: outreach messages adapted for cultural context — understanding that cancer screening messaging may need to address fatalism beliefs in some cultural communities, that vaccination outreach may need to address specific concerns prevalent in certain populations, and that wellness visit messaging may need to emphasise family health rather than individual health for cultures where family decision-making is the norm. Channel optimisation: messages delivered through the patient's preferred channel (patient portal, SMS, email, or mail) in their preferred language — the GPT generating format-appropriate content for each channel. Clinical documentation assistance (weeks 10-14). Physician copilot: GPT integrated into the Epic documentation workflow — generating draft progress notes from structured encounter data (vital signs, chief complaint, examination findings, orders placed). The drafts: formatted to the health system's documentation standards with the physician reviewing and editing rather than writing from scratch. Referral letters: GPT generating referral letters with relevant clinical history pulled from the patient's EMR record — the physician reviewing and signing rather than dictating or typing. Prior authorisation: GPT drafting prior authorisation justifications — pulling clinical evidence from the patient record and matching against payer-specific requirements. The drafts: significantly reducing the 25-45 minutes staff spent preparing each prior auth request.

The GPT integration deployed across all 11 facilities within 16 weeks. Patient communication quality: after-visit summary readability: from grade 12 average to grade 6.8 — meeting the recommended health literacy standard. Patient comprehension: from 42% to 74% of patients reporting full understanding of their care instructions (78% improvement). Spanish-language quality: patient satisfaction with Spanish communication improved 56% — native-language generation receiving significantly higher quality ratings than the previous translation-based approach. Care instruction adherence: improved 28% among patients receiving GPT-generated summaries vs. the previous standard summaries. Outreach effectiveness: preventive care outreach response rate: from 12% (English) / 8% (Spanish) to 24% (English) / 22% (Spanish) — personalised, culturally adapted messaging nearly doubling response rates and closing the language gap. Cancer screening: mammography screening completion increased 18% among the target population — driven by culturally adapted outreach that addressed specific barriers. Vaccination: flu vaccination rates improved 22% among Spanish-speaking patients — the personalised outreach closing the vaccination gap between English and Spanish-speaking populations. Documentation efficiency: physician documentation time: from 2.1 hours daily to 1.3 hours (38% reduction) — GPT-generated drafts enabling review-and-edit rather than create-from-scratch. Prior authorisation preparation: from 25-45 minutes to 8-12 minutes — GPT pulling clinical evidence and drafting justifications that staff reviewed and submitted. Referral letter generation: from 12-18 minutes to 3-5 minutes — GPT drafting letters with relevant clinical history automatically. Physician satisfaction: documentation satisfaction improved 42% — the reduction in documentation burden directly addressing the primary source of professional dissatisfaction. Financial results: care adherence improvement (reduced readmissions, fewer complications): $3.8M/year. Preventive care improvement (increased screening revenue + reduced downstream costs): $2.4M/year. Physician productivity (documentation time recovery): $1.6M/year. Prior authorisation efficiency: $480K/year. Total annual impact: $8.28M/year. GPT integration investment: $420,000. Annual operations (Azure OpenAI, monitoring, validation): $168,000/year. First-year ROI: 1,871%.