research

Physiological Signal Learning

Research question. Physiological signals are long, noisy, and vary between individuals. Can a model learn directly from raw or minimally processed signals, and how does it compare with established handcrafted features?

Signals and learning problems.

• Surface EMG (three forearm channels): classification of ten hand and finger gestures, seven of them individual-finger gestures.
• ECG-derived RR intervals: classification of four age groups in healthy adults.
• RR intervals: detection of acute cognitive stress in major depressive disorder, panic disorder, and healthy controls, from 5-minute and 1-minute windows.

Models compared. For EMG, an artificial neural network against SVM, random forest, and logistic regression on time-domain features. For age, a modified ResNet on raw RR intervals against a heart-rate-variability (HRV) feature model and a combined model. For stress, a one-dimensional ResNet-34 on raw RR intervals, trained per diagnostic group and on pooled data.

Data and validation. Ten subjects in a prospective EMG experiment (personalized classifiers); RR recordings from 1,093 healthy adults; 147 adults with up to five laboratory visits over 12 weeks, each with a resting baseline and a mental-arithmetic stressor. In the age study the raw-RR model performed comparably to the HRV model, but no model predicted every age group reliably — a pilot result rather than a usable age estimator.

My role.

• EMG — co-first author: study protocol and IRB approval, subject recruitment and experiments, real-time filtering, and classifier development.
• Age prediction — first author.
• Stress detection — co-first author: data curation, methodology, software, formal analysis, and validation.

Papers and code.

• Lee†, Min† & Byun*. Sensors, 2022. doi:10.3390/s22010225 · code
• Lee† & Byun*. Applied Sciences, 2023. doi:10.3390/app13052932 · code
• Lee†, Cho†, Kim, Jeon* & Byun*. Frontiers in Psychiatry, 2025. doi:10.3389/fpsyt.2025.1672260 · code

Clinical Prediction

Research question. Can a model answer a specific clinical question — does this patient have asthma, is this inpatient developing sepsis — better than the rule clinicians already use, and does the answer hold on patients the model has not seen: later referrals, or a prospective cohort?

Methods.

• Asthma: five machine learning models (logistic regression, SVM, random forest, XGBoost, neural network) that read the full spirometric response during a methacholine challenge test (FEV1, FVC, FEF25–75), compared with the conventional PC20 ≤ 16 mg/mL criterion; Shapley values reviewed with clinicians.
• Sepsis: a deep learning score computed from electronic medical records (VC-SEPS), compared prospectively with conventional scoring systems, including risk stratification at admission and lead time before an operational sepsis definition was met.

Data and validation. 1,501 patients referred for methacholine challenge at Samsung Medical Center (2015–2020), split by referral date so that the test set is later than the training set. 6,455 hospitalised patients (325 with sepsis) in a prospective observational study at Keimyung University Dongsan Hospital.

My role.

• Asthma — co-first author: cohort definition with clinicians, data extraction from the clinical data warehouse, model development, and Shapley-value review.
• Sepsis — co-first author.

Papers and code.

• Kang†, Lee†, Byun, Lee, Choi & Lee*. Allergy, Asthma & Immunology Research, 2024. doi:10.4168/aair.2024.16.1.42 · code
• Kim†, Lee†, et al. BMJ Health & Care Informatics, 2025. doi:10.1136/bmjhci-2024-101353

Related co-authored work on deterioration and sepsis prediction: prospective external validation of a general-ward early-warning system (Acute and Critical Care, 2025), real-time deterioration prediction in the emergency department (Scientific Reports, 2024), and pneumonia-associated sepsis (Journal of Clinical Medicine, 2026). See publications.

Real-World Evaluation

Research question. A deployed early-warning model re-predicts every time new measurements arrive, but models are usually compared on a single prediction per patient. Does that single-point performance describe how a model behaves when it is read repeatedly during a stay, and does it carry over to other hospitals?

Methods.

• A labeling algorithm that defines AKI onset in electronic health records, as the basis for early-prediction targets.
• Three deep learning architectures (LSTM-Attention, Masked CNN, ITE-Transformer) and two baselines (XGBoost, logistic regression) at 0-, 48-, and 72-hour horizons.
• An online simulation that replays each stay and issues predictions every 12 hours before onset, evaluated for clinical faithfulness (whether discrimination improves consistently as onset approaches, tested with Mann–Kendall trends) and for alert burden (number needed to evaluate).

Data and validation. 157,323 admissions from three cohorts: NHIS Ilsan Hospital (development), Chuncheon Sacred Heart Hospital, and MIMIC-IV (external validation). Deep learning models discriminated well externally, but the model with the highest single-point AUROC had the least favourable alert burden in simulation, while a model with lower AUROC had the most favourable one.

My role. Co-first author of the npj Digital Medicine study — conceptualization, methodology, software, data curation, formal analysis, validation, and the original draft. First author of the labeling-algorithm paper. At AITRICS, I lead product development of the AKI model toward approval as a medical device by Korea’s MFDS.

Papers.

• Lee†, Yoon†, Lim, Lee* & Lee*. npj Digital Medicine, 2026. doi:10.1038/s41746-026-02722-2
• Lee†, Hahn, Lim, Lee & Lee*. Development of labeling algorithm for early prediction of acute kidney injury. MIE 2025 (Studies in Health Technology and Informatics). doi:10.3233/SHTI250415
• Conference abstracts at AKI & CRRT 2026 on online simulation and prediction horizons, and on nursing-note embeddings from a large language model — see publications.

Related co-authored proceedings on evaluating early-prediction models: time-varying alert thresholds, and activity monitoring as an alternative to the ROC curve (STC 2024).

Code for this work is not publicly available.

Longitudinal Pulmonary Function

Status. Ongoing doctoral research. Presented as a late-breaking poster at the ERS International Congress 2026; not yet published as a paper.

Research question. In health screening, spirometry is repeated over many years, yet each result is usually read on its own. Does the trajectory of repeated tests carry information about future airflow obstruction beyond the latest value, and is a nonlinear sequence model needed to capture it?

Methods. A landmark design: at three years after the first test, predict incident airflow obstruction (FEV1/FVC < 0.70) over the next five years using only earlier tests. Three readings of the same history are compared — the latest value, a linear trajectory summary (slope, variability, test count, span), and a sequence model that reads every measurement in order.

Data and validation. 557,503 screening spirometry tests from 221,903 adults with normal airflow at their first test, from Samsung Medical Center. Internal validation, with sensitivity analyses that exclude outcomes within one or two years of the landmark, require two consecutive obstructed tests, or restrict to a higher baseline FEV1/FVC. The data are retrospective and single-centre, without smoking or diagnosis labels; clinically performed pulmonary function tests and a multicenter external validation are the planned next steps.

My role. First author, with Dayoon Lee, Prof. Jin-Young Lee, and Prof. Byung-Jae Lee (doctoral advisor; corresponding author).

Links. Poster (PDF) · Conference notes

Evaluating Continuously Read Clinical Models

Status. Manuscripts under review and a preprint; none of this work is published in a journal yet.

Research question. For a clinical model that is read continuously — every hour on a ward, or every minute in the operating room — which evaluation choices change the conclusion about which model is best?

Ongoing work.

• Minute-level intraoperative hypotension monitoring — whether commonly used metrics agree when ranking a model that is read continuously. First author; manuscript under review.
• Transformer-based cardiac arrest prediction for general-ward patients — development and multicenter external validation. Co-first author; a development preprint is on medRxiv and the multicenter manuscript is under review.
• Nursing notes for AKI prediction — adding large-language-model embeddings of nursing notes to time-series AKI models, led by Donghwee Yoon. Co-author; presented at AKI & CRRT 2026.

See Preprints and Under Review for the current status of each manuscript.

Multimodal Emotion Recognition from Speech and Text

Emotion recognition models that combine audio and text sequences, including a self-supervised contrastive loss. Published as short papers in the Korea Computer Congress proceedings (not journal articles):

• Nam† & Lee*. Development and validation of emotion recognition model using audio and text sequences: two heads are better than one. Korea Computer Congress, 2022. Corresponding author.
• Nam†, Lee†, Choi & Ha*. Self-supervised contrastive loss for multimodal emotion recognition. Korea Computer Congress, 2023. Co-first author.