# Research ## 4D flow risk assessment for anomalous coronary origin in children (2025–2028) Page: https://hyeongjunlee.me/research/#anomalous-origin In this device project, 4D flow grades the risk of an anomalous aortic origin of a coronary artery in children and adolescents. Outputs: - Funded project: 4D Flow-Based Diagnostic Device for Dynamic Risk Assessment of Anomalous Origin of the Coronary Artery in Children and Adolescents. Ministry of Health and Welfare / KHIDI, Pediatric Disease R&D Program · Apr. 2025 – Dec. 2028 (participating researcher) ## Neural operator for patient-specific coronary flow (2024 –) Page: https://hyeongjunlee.me/research/#pino A physics-informed neural operator trained on CFD returns a vessel's pressure and velocity field in 1.78 s instead of a 35-minute run. Outputs: - Paper: Patient-Specific Coronary Flow Field Prediction Using Physics-Informed Neural Operators. Proc. ASME 2026 Design of Medical Devices Conference (DMD2026), Minneapolis, MN, USA · V001T02A010 · 2026 · First author · doi:[10.1115/DMD2026-1091](https://doi.org/10.1115/DMD2026-1091) - Patent: Patient-specific coronary flow prediction method based on physics-informed neural operators. Pending, filed 2026 · KR 10-2026-0003225 - Talk: Patient-Specific Coronary Flow Field Prediction Using Physics-Informed Neural Operators. ASME Design of Medical Devices Conference (DMD2026), Minneapolis, MN, USA, 2026 · poster · presenter - Talk: Patient-Specific Coronary Flow Field Prediction Using Physics-Informed Neural Operators. 18th Asian Congress of Fluid Mechanics (ACFM), Seoul, Korea, 2025 · oral · presenter - Talk: Modeling Coronary Artery Hemodynamics: Exploring DCNN Surrogate Models in Preliminary Research. Biomedical Engineering Society for Circulation (BESCO) Summer Meeting, Daegu, 2023 · oral · presenter `CFD up to 35 min on 64 CPU cores · PINO 1.78 s on one L40S` Artifact: a real patient vessel (PINO-DEMO-01, 8 649 surface nodes, 17 294 triangles, 296 streamlines) painted with its CFD pressure field and the field the PINO returned for the same geometry, with a CFD | PINO wipe, a Pressure | Error | Speed toggle and the Pd/Pa curve along the vessel. FFR is read on the probe planes at s = 5.8 mm and 66.0 mm; the CFD and PINO readings are 0.006 apart. Colorbars: pressure 10.7–13.0 kPa, one range for CFD and PINO (ticks 10.7 / 11.3 / 11.9 / 12.4 / 13.0 kPa); error PINO − CFD on a fixed symmetric scale of ±1.0 kPa (ticks −1.0 / −0.5 / 0 / +0.5 / +1.0 kPa); near-wall speed 0–0.14 m/s (ticks 0.00 / 0.03 / 0.07 / 0.10 / 0.14 m/s). Plate notes: error, “PINO minus CFD at each node. Here −0.2 to +0.6 kPa, largest at the throat.”; speed, “Speed one node inside the no-slip wall; inlet and outlet caps masked.” The current PINO is a geometry-informed neural operator (GINO backbone) that takes the vessel surface and returns pressure and velocity at every volume node. Its loss carries a pressure-only Navier–Stokes residual at the prescribed and at counterfactual inlet flows. It was trained and evaluated by 10-fold cross-validation over 85 patients; this case is predicted by the fold model that never saw it. Inference over ~80k volume nodes takes 1.78 s on one NVIDIA L40S; the reference CFD solve takes up to 35 min on 64 CPU cores (owner, 2026-10-16, an upper bound). Metrics: current model, 85 cases, 10-fold CV, all volume nodes of each case. | Metric | This case | Cohort median, n = 85 | | --- | --- | --- | | Pressure R² | 0.963 | 0.898 | | Pressure MAE | 0.1 kPa | 0.2 kPa | | Speed R² | 0.846 | 0.756 | | FFR, Pd/Pa | CFD 0.830 · PINO 0.824 | MAE vs CFD 0.038 | A well-predicted case, not a typical one: its pressure R² ranks 13 of 85 in the cohort. ## Dynamic FFR from diffusion-synthesized 4D CT (2024–2026) Page: https://hyeongjunlee.me/research/#ct4d Two acquired CT phases become a full cardiac cycle; each synthesized phase is meshed and solved, so FFR is read across the cycle at 90.5 % accuracy against invasive FFR. Outputs: - Paper: Synthesis of coronary 4D CT image by denoising diffusion probabilistic model. Computer Methods and Programs in Biomedicine · 282, 109382 · 2026 · Co-author · doi:[10.1016/j.cmpb.2026.109382](https://doi.org/10.1016/j.cmpb.2026.109382) - Paper: Deriving Dynamic Fractional Flow Reserve From 4D Coronary CT Synthesized by a Diffusion Model. Proc. ASME 2026 Design of Medical Devices Conference (DMD2026), Minneapolis, MN, USA · V001T02A011 · 2026 · Co-author · doi:[10.1115/DMD2026-1092](https://doi.org/10.1115/DMD2026-1092) - Patent: Deep learning-based multiphase coronary CT interpolation system and method. Pending, filed 2025 · US 19/397,163 - Patent: Deep learning-based multiphase coronary CT interpolation system and method. Pending, filed 2025 · KR 10-2025-0149607 - Patent: Deep learning-based multiphase coronary CT interpolation method. Pending, filed 2024 · KR 10-2024-0168733 - Funded project: 4D Coronary Flow Visualization from CT in Real Time (4D CT-FFR). Ministry of Health and Welfare / KHIDI, Clinical-Needs Translational Research · Apr. 2024 – Dec. 2026 (participating researcher) - Talk: Dynamic Fractional Flow Reserve from 4D Coronary CT Synthesized by a Diffusion Model. ASME Design of Medical Devices Conference (DMD2026), Minneapolis, MN, USA, 2026 · poster · co-author - Talk: Time-resolved Coronary Flow Visualization from Single-phase CT. Biomedical Engineering Society for Circulation (BESCO) Summer Meeting, Gwangmyeong, 2025 · poster · co-author Artifact: Case 1 from the 4D CT paper (J1, Fig. 8) through one cardiac cycle, pressure and wall shear stress at each of 10 phases; P0 and P5 were acquired by CT, the other 8 synthesized, and CFD solved every one. Colour scales (the paper's renders): Pressure 70–130 mmHg (9.3–17.3 kPa); Wall shear stress 0–5 Pa. A denoising diffusion deformable model synthesizes a full-cycle 4D coronary CT from two acquired phases; the sequence is then meshed and solved, so FFR can be followed across the cardiac cycle instead of read from one static phase. | Metric (reported in the paper) | Value | | --- | --- | | Phases | 10 (P0 and P5 acquired) | | 4D CT-FFR accuracy against invasive FFR | 90.5 % | | Synthesis PSNR | 32.01 | | Synthesis SSIM | 0.937 | | Synthesis Dice | 0.973 | ## Fluid-bolus prediction from smartwatch PPG (2024 –) Page: https://hyeongjunlee.me/research/#bolus Thirty CNNs read PPG spectrograms and a weighted average predicts an intravenous fluid bolus within 24 hours of emergency-department arrival (AUC 0.895). Outputs: - Paper: Multi-Model Deep-Learning Ensemble for Predicting 24-Hour Fluid Bolus Administration from Continuous Photoplethysmography Spectrogram Analysis. Engineering Applications of Artificial Intelligence · First author · First revision under review Artifact: a 60 s smartwatch PPG recording as the network sees it, an STFT spectrogram (129 × 360, 0–5 Hz shown) and a CWT strip (128 × 600), from the development cohort, taken before the first bolus. Green-channel PPG at 100 Hz is band-passed, normalized in short windows and cut into overlapping segments; each segment becomes an STFT or CWT image. Six CNN configurations across four backbones, each kept as five fold checkpoints, make 30 networks. Segment scores are averaged per patient, rank-normalized within each fold and combined by a non-negative weighted average. Validation is 5-fold stratified cross-validation with out-of-fold, patient-level scores. Fluid-bolus prediction at patient level, from M1, manuscript under revision (EAAI). | Set | AUC [95 % CI] | n | | --- | --- | --- | | Development, out-of-fold | 0.895 [0.839–0.940] | 195 | | Same-center temporal set | 0.827 [0.746–0.898] | 141 | The development figure is internally selected (ensemble weights and cutoff were chosen on those predictions); the temporal set was scored with the model, weights and cutoff unchanged. ## Patient-specific coronary artery disease risk analysis software (2024–2026) Page: https://hyeongjunlee.me/research/#cad-software In this SME-funded software project, a patient's CT becomes hemodynamic risk indices. Outputs: - Funded project: Patient-Specific Coronary Artery Disease Risk Analysis Software. Ministry of SMEs and Startups, Startup Growth Technology Development Program (TIPS) · Jul. 2024 – Jun. 2026 (participating researcher) ## Blood viscosity from smartwatch PPG (2023–2025) Page: https://hyeongjunlee.me/research/#viscosity A physics-constrained CNN–LSTM with a Carreau–Yasuda prior estimates a patient's non-Newtonian blood viscosity from PPG and basic biometrics; the paper received the Merit Academic Paper Award 2026. Outputs: - Paper: A Physics-Integrated Deep Learning Approach for Patient-Specific Non-Newtonian Blood Viscosity Assessment using PPG. Computer Methods and Programs in Biomedicine · 265, 108740 · 2025 · First author · doi:[10.1016/j.cmpb.2025.108740](https://doi.org/10.1016/j.cmpb.2025.108740) - Patent: Device and method for predicting viscosity using AI in a viscosity prediction system. Pending, filed 2025 · PCT/KR2025/002462 - Patent: Device and method for predicting viscosity using AI in a viscosity prediction system. Pending, filed 2025 · KR 10-2025-0021284 - Patent: Non-Newtonian fluid viscosity modeling of patient blood using wearable device-based PPG and biometric information. Pending, filed 2024 · KR 10-2024-0026709 - Patent: Systolic and diastolic viscosity prediction algorithm using wearable device-based PPG and biometric information. Pending, filed 2024 · KR 10-2024-0028201 - Talk: A Physics-Integrated Deep Learning Approach for Patient-Specific Non-Newtonian Blood Viscosity Assessment using PPG. Biomedical Engineering Society for Circulation (BESCO) Summer Meeting, Gwangmyeong, 2025 · poster · presenter - Talk: AI-based Hemorheology Prediction with Patient-Specific Biometric Boundary Conditions. 26th International Congress of Theoretical and Applied Mechanics (ICTAM), Daegu, Korea, 2024 · short oral & poster · presenter - Talk: Arterial Blood Pressure and Blood Viscosity Estimation from AI-based PPG Analysis. Biomedical Engineering Society for Circulation (BESCO) Winter Meeting, Seongnam, 2024 · poster · co-author - Talk: Blood viscosity estimation from wearable device using AI-based PPG analysis. Biomedical Engineering Society for Circulation (BESCO) Summer Meeting, Daejeon, 2024 · poster · co-author ## Wearable screening for glucose, diabetes and glaucoma (2023–2024) Page: https://hyeongjunlee.me/research/#wearable-screening PPG and contactless biosignal models estimate blood glucose and diabetes risk and flag glaucoma suspects; they are filed as patents. Outputs: - Patent: Method for predicting blood glucose and diabetes using PPG signals and biometric information. Pending, filed 2024 · KR 10-2024-0147541 - Patent: Glucose and diabetes prediction algorithm using wearable device-based PPG and biometric information. Pending, filed 2023 · KR 10-2023-0144347 - Patent: Wearable device for monitoring glaucoma suspect and method for monitoring glaucoma suspect. Pending, filed 2024 · PCT/KR2024/012365 - Patent: Wearable device for monitoring glaucoma suspect and method for monitoring glaucoma suspect. Pending, filed 2023 · KR 10-2023-0118582 - Patent: Glaucoma diagnosis method and system based on contactless biosignals. Pending, filed 2023 · KR 10-2023-0078883 - Talk: Unlocking Predictive Health Outcomes with Biometric Data. Korean Society of Mechanical Engineers (KSME) Conference, Songdo, 2023 · oral · presenter ## Real-time precision medicine platform on smart hemodynamic indices (2022–2029) Page: https://hyeongjunlee.me/research/#platform This seven-year NRF Leading Research Center project studies hemodynamic indices for real-time precision medicine. Outputs: - Funded project: Real-Time Precision Medicine Platform Based on Smart Hemodynamic Indices. National Research Foundation of Korea (NRF), Leading Research Center Program · Jun. 2022 – Feb. 2029 (participating researcher) ## Variable negative-pressure tunnel for respiratory infection control (2022–2024) Page: https://hyeongjunlee.me/research/#tunnel This KHIDI project develops a variable negative-pressure tunnel for respiratory infection control. Outputs: - Funded project: Variable Negative-Pressure Tunnel for Respiratory Infection Control. Ministry of Health and Welfare / KHIDI, Infectious Disease Medical Safety Technology Program · Apr. 2022 – Dec. 2024 (participating researcher) ## AI-based non-invasive urodynamics test (2022) Page: https://hyeongjunlee.me/research/#urodynamics A registered patent covers a non-invasive urodynamics test method and apparatus. Outputs: - Patent: AI-based noninvasive urodynamics test method and apparatus. Registered Dec. 2025 · KR 10-2022-0085583 (reg. 10-2902500) ## FFR prediction in the gray zone (2020–2022) Page: https://hyeongjunlee.me/research/#gray-zone An AI algorithm combines hemodynamic features, synthetic vessel models and biometrics to call the FFR gray zone; it is published and patented, with a quantum-ML follow-up. Outputs: - Paper: Optimization of FFR prediction algorithm for gray zone by hemodynamic features with synthetic model and biometric data. Computer Methods and Programs in Biomedicine · 220, 106827 · 2022 · Co-first author · doi:[10.1016/j.cmpb.2022.106827](https://doi.org/10.1016/j.cmpb.2022.106827) - Patent: Optimization system and method of AI algorithm for prediction coronary artery lesions based on FFR. Registered Nov. 2025 · KR 10-2022-0030019 (reg. 10-2882193) - Patent: Optimization system and method of AI algorithm for prediction coronary artery lesions based on FFR. Pending, filed 2022 · US 17/820,819 - Talk: Optimization of Artificial Intelligence Algorithms for FFR Prediction in Gray Zone. 10th International Conference of Manufacturing Technology Engineers (ICMTE), Gyeonggi-do, Korea, 2022 · oral · presenter - Talk: Artificial Intelligence Algorithms for FFR Prediction in Gray Zone by Single-view Angiography. Biomedical Engineering Society for Circulation (BESCO) Winter Meeting, Seoul, 2022 · short oral & poster; Best Paper Award · presenter - Talk: FFR Prediction with Quantum Machine Learning: Accurate and Non-invasive Diagnosis with Limited Features. Biomedical Engineering Society for Circulation (BESCO) Winter Meeting, Seongnam, 2024 · poster · co-author ## CFD-based CT-FFR with the lattice Boltzmann method (2020–2022) Page: https://hyeongjunlee.me/research/#lbm-ctffr Image-based, patient-specific coronary CFD runs on a lattice Boltzmann solver after automatic 3D segmentation. Outputs: - Talk: Estimating CFD-based CT FFR using lattice Boltzmann method – 3D geometry auto segmentation and novel patient-specific computation. ESCHM-ISCH-ISB 2021, Fukuoka, Japan, 2021 · oral · presenter - Talk: Autonomous Cardiovascular Hemodynamics Analysis using Patient-Specific CFD Simulation with Unsupervised Learning-based Coronary Artery Segmentation. Biomedical Engineering Society for Circulation (BESCO) Summer Meeting, Seoul, 2021 · oral · co-author ## High-speed point-of-care FFR simulator (2020–2022) Page: https://hyeongjunlee.me/research/#ffr-simulator This NRF project works toward a fast FFR simulator for diagnosis and treatment at the point of care. Outputs: - Funded project: High-Speed Point-of-Care FFR Simulator for Cardiovascular Diagnosis and Treatment. NRF, Bio & Medical Technology Development Program · Feb. 2020 – May 2022 (participating researcher) ## Tricuspid regurgitation membrane device (2020–2025) Page: https://hyeongjunlee.me/research/#tricuspid CFD optimizes a carved membrane device for tricuspid regurgitation; the goal is a patient-specific implant made with 3D heart printing. Outputs: - Paper: Optimization of tricuspid membrane mechanism for effectiveness and leaflet longevity through hemodynamic analysis. Engineering Applications of Computational Fluid Mechanics · 16(1), 1587–1600 · 2022 · Co-author · doi:[10.1080/19942060.2022.2104929](https://doi.org/10.1080/19942060.2022.2104929) - Funded project: Patient-Specific Implantable Device for Tricuspid Regurgitation Using 3D Heart Printing and 4D-CFD Simulation. Ministry of Health and Welfare / Korea Medical Device Development Fund · Sep. 2020 – Dec. 2025 (participating researcher) - Talk: Carved membrane design for tricuspid valve regurgitation treatment. Biomedical Engineering Society for Circulation (BESCO) Summer Meeting, Daejeon, 2024 · poster · co-author --- Open to collaboration, positions and paper requests. [Email](mailto:hjunlee96@gmail.com) · [CV (PDF)](https://hyeongjunlee.me/uploads/resume.pdf) · [LinkedIn](https://www.linkedin.com/in/hyeongjunlee/) · [Google Scholar](https://scholar.google.com/citations?user=1NxM3T0AAAAJ&hl=en) · [ORCID](https://orcid.org/0009-0001-5691-3787) Updated 2026-10-21 · hyeongjunlee.me