Overview

Patients with chronic disease in Singapore spend just 15 minutes a year with their doctor and 4,300 hours managing their health alone, a gap that drives 56% medication non-adherence and SGD 2.5B in preventable A&E costs. Med-SEAL is a SEA-LION-empowered, FHIR-native AI companion that fills that gap with a 24/7 multilingual chat agent, proactive medication nudges, culturally aware dietary coaching, and a 30-day pre-visit brief for clinicians. Built on Singapore’s national LLM with certified adversarial robustness, it turns chronic care from 15 minutes of reactive encounters into continuous, proactive patient empowerment.

Deployment

Med-SEAL is built as a Mixture-of-Agents (MoA) medical AI on top of Singapore’s SEA-LION National LLM, with every patient interaction grounded in a FHIR-native clinical knowledge graph rather than a flat health record. When a patient asks a question, three specialist agents work in parallel: a multilingual reasoning agent (SEA-LION-v4) handles the conversation in English, Mandarin, Malay, or Tamil; a medical reasoning agent (Med-SEAL-V1, our adversarially-trained clinical VLM) checks drug interactions and clinical logic; and a knowledge graph retriever pulls the patient’s real conditions, medications, and lab trends from the EMR. A SEA-LION aggregator then resolves any conflicts, enforces safety policy, and returns a single grounded answer, wrapped at both input and output by a SEA-Guard safety layer.

Key components:

• Med-SEAL Companion mobile app with 24/7 chat, medication tracker, vitals dashboard, appointment booking, and SEA-culturally-aware dietary coaching (hawker food, festive meals).
• 7-section pre-visit brief auto-generated 24 hours before each appointment, giving clinicians 30 days of adherence, biometric, and patient-reported outcome data in one place.
• Tiered nudge engine that sends gentle reminders for missed doses (low), flags clinicians next day for concerning trends (medium), and triggers immediate alerts for dangerous readings (high).
• FHIR R4 integration with OpenEMR, Medplum, and Epic on FHIR, making Med-SEAL interoperable with any hospital system in Singapore.

What makes it unique: Med-SEAL is the first medical VLM in Southeast Asia to combine adversarial training during GRPO reinforcement learning (achieving 0.74% attack success rate, 40 times lower than the next-best medical VLM) with certified robustness guarantees via randomized smoothing. Built in partnership with NUS, Synapxe, IMDA, and MOH Singapore, and aligned with MOH AIHGIe 2.0 guidelines, it is designed to be safe by construction, not safe by filter.

Conclusion

Med-SEAL has been validated across three independent dimensions. On clinical robustness, Med-SEAL-V1 achieves 0.74% attack success rate on OmniMedVQA under PGD-20 adversarial attack, 40 times lower than the next-best medical VLM. On safety, independent red-teaming using AI Verify 2.0 (11/11 principles), Microsoft PyRIT (26/26 OWASP probes), and NVIDIA Garak (33/33 probes) returned zero breaches across 230 attack scenarios. On quality, LLM-as-judge evaluation with DeepEval v3.9.6 scored the system at 0.986 faithfulness, 0.940 answer relevancy, and 0.933 clinical safety, with zero bias and zero toxicity detected.

Next steps are focused on real-world deployment. We are seeking clinical pilot partners among Singapore’s polyclinics and private hospital groups to validate Med-SEAL with 500 to 1,000 chronic disease patients over a six-month intervention, measuring medication adherence (PDC), biometric control (HbA1c, blood pressure), and clinician time saved per consultation. In parallel, we are pursuing MOH AIHGIe 2.0 certification, deeper integration with Epic on FHIR and Synapxe’s systems, and expansion to Indonesia and Malaysia in partnership with AI Singapore’s SEA-LION programme.

Our vision is simple: every chronic disease patient in Southeast Asia deserves a trustworthy, multilingual, culturally aware health companion in those 4,300 hours between clinic visits. Med-SEAL is how we get there, safely, starting with Singapore.

Team Med-SEAL is a multidisciplinary team of six from the National University of Singapore, spanning business analytics, biomedical informatics, and AI safety research. We came together around a shared conviction: Southeast Asian patients deserve healthcare AI that speaks their languages, understands their food, respects their privacy, and refuses to fail silently.

Yogi (A.A. Gde Yogi Pramana) is an MComp in AI student at NUS Computing & AI Safety and Security Researcher at the NUS Artificial Intelligence Institute, supervised by Prof. Mohan Kankanhalli. His research on adversarial robustness and certified defenses for vision-language models forms the technical backbone of Med-SEAL-V1. He leads the team’s AI architecture and trustworthy-AI evaluation.

Andini Sumardi, Carlo Ida, Lin Frankie Yang, and Oishika Saha are MSc candidates in Business Analytics at NUS. Between them, they lead Med-SEAL’s product design, clinical workflow mapping, market and economic analysis, and the Med-SEAL Companion mobile application.

Yuliana Patangke is an MSc candidate in Biomedical Informatics at NUS, leading FHIR R4 data modelling, clinical knowledge graph construction, and integration with OpenEMR, Medplum, and Epic on FHIR.

What motivates us is personal as much as professional. Several team members come from ASEAN countries where a parent, grandparent, or neighbor manages a chronic condition without the language, literacy, or access that global health AI assumes. We are building Med-SEAL for them first, and for the millions like them across Singapore, Indonesia, Malaysia, and the wider ASEAN region.