Large Multimodal Model (LMM) sebagai Clinical Decision Support System (CDSS) dalam Sistem Rujukan Kesehatan Gigi dan Mulut

Achmad Zam Zam Aghasy, Prof. dr. Hari Kusnanto Josef, SU., Sp.KKLP, Dr.PH

2026 | Disertasi | S3 Kedokteran Umum

Latar Belakang: Sistem rujukan kesehatan gigi di Indonesia mengalami tingkat ketidaktepatan tinggi (52,2-58,2%) antara Fasilitas Kesehatan Tingkat Pertama (FKTP) dan Lanjutan (FKTL) akibat kurangnya pemahaman diagnosis serta kode ICD-10 dan ICD-9-CM. Tujuan: Penelitian ini mengevaluasi potensi Large Multimodal Model (LMM) sebagai Clinical Decision Support System (CDSS) untuk mengatasi kesenjangan tersebut dengan mengukur performa berbagai model LMM. Metode:  Menggunakan desain mixed-methods sekuensial eksplanatori dalam tiga tahap. Tahap 1 mengevaluasi 11 model LMM (5 API hosted dan 6 self-hosted) dengan 140 soal Uji Kompetensi Mahasiswa Program Profesi Dokter Gigi (UKMP2DG) dari 14 departemen. Tahap 2 merupakan studi Randomized Controlled Trial (RCT) open-label dengan dua kelompok paralel yang melibatkan 62 dokter gigi FKTP di Daerah Istimewa Yogyakarta, membandingkan kelompok AI (n = 31) dan Non-AI (n = 31) dalam mendiagnosis 10 kasus klinis. Kelompok AI menggunakan model Claude (Claude Opus 4-20250514) yang diimplementasikan pada kondisi baseline dengan pendekatan role dan simulation prompting; risiko halusinasi dimitigasi melalui penilaian outcome objektif berbasis kunci jawaban dan desain human-in-the-loop di mana dokter tetap memvalidasi keluaran AI. Tahap 3 mengeksplorasi penerimaan teknologi berdasarkan kerangka kerja UTAUT2 dan prinsip etika AI WHO melalui Focus Group Discussion (FGD) dengan 8 partisipan. Analisis data menggunakan uji Chi-Square, Kruskal-Wallis, Mann-Whitney U, MANOVA, dan analisis terarah. Hasil: Tahap 1 menunjukkan Claude mencapai akurasi tertinggi (71,4%), dengan model API hosted secara signifikan mengungguli model self-hosted (67,9% vs 43,0%; ?² = 95,2; p < 0 xss=removed xss=removed>Kesimpulan: Model LMM API hosted, khususnya Claude, menunjukkan performa superior dalam menjawab soal kompetensi kedokteran gigi dibandingkan model self-hosted. CDSS berbasis LMM terbukti efektif meningkatkan ketepatan diagnosis ICD-10 dan prosedur medis ICD-9-CM dokter gigi FKTP secara substansial tanpa memperpanjang waktu diagnosis. Implementasi CDSS memerlukan dukungan infrastruktur yang memadai dan perhatian terhadap aspek etika AI.

Background: The dental referral system in Indonesia faces high inaccuracy rates (52.2–58.2%) between Primary Healthcare Facilities (FKTP) and Advanced Healthcare Facilities (FKTL) due to suboptimal diagnostic understanding and limited proficiency in ICD-10 and ICD-9-CM coding. Objective: This study evaluated the potential of Large Multimodal Models (LMM) as a Clinical Decision Support System (CDSS) to bridge this gap by measuring the performance of various LMMs. Methods: This study employed an explanatory sequential mixed-methods design across three phases. Phase 1 evaluated 11 LMMs (5 paid APIs and 6 self-hosted models) using 140 questions from the Indonesian Dental Student Competency Examination (UKMP2DG) spanning 14 departments. Phase 2 was an open-label, parallel-group Randomized Controlled Trial (RCT) involving 62 primary care dentists in Yogyakarta, comparing AI (n = 31) and Non-AI (n = 31) groups in diagnosing 10 clinical cases. The AI group deployed the Claude model (Claude Opus 4-20250514) at baseline using role and simulation prompting. Hallucination risks were mitigated via answer key-based objective assessment and human-in-the-loop validation by doctors. Phase 3 explored technology acceptance based on the UTAUT2 framework and WHO AI ethical principles through a Focus Group Discussion (FGD) with 8 participants. Data analysis utilized Chi-Square, Kruskal-Wallis, Mann-Whitney U, MANOVA, and directed content analysis. Results: Phase 1 demonstrated Claude achieved the highest accuracy (71.4%), with commercial API models significantly outperforming self-hosted models (67.9% vs 43.0%; ?² = 95.2; p < 0 xss=removed xss=removed>Conclusions: Commercial API-based LMM models, particularly Claude, demonstrated superior performance in answering dental competency questions compared to self-hosted models. LMM-based CDSS proved effective in substantially improving ICD-10 diagnostic accuracy and ICD-9-CM medical procedure accuracy among FKTP dentists without prolonging diagnosis time. CDSS implementation requires adequate infrastructure support and attention to AI ethics aspects.

Kata Kunci : Large Multimodal Model, Clinical Decision Support System, Sistem Rujukan, Kesehatan Gigi dan Mulut, UTAUT2