Organizations: Institute of Medical Technology, Peking University Health Science Center, Beijing, China · National Institute of Health Data Science, Peking University, Beijing, China · Department of Cardiology, Peking University People’s Hospital, Beijing, China · School of Intelligence Science and Technology, Peking University, Beijing, China · Heart Voice Medical Technology, Hefei, China · Renal Division, Department of Medicine, Peking University First Hospital, Beijing, China · Institute of Nephrology, Key Laboratory of Renal Disease, Ministry of Health of China, and Key Laboratory of Chronic Kidney Disease Prevention and Treatment (Peking University), Ministry of Education, Beijing, China · Research Units of Diagnosis and Treatment of Immune-mediated Kidney Diseases, Chinese Academy of Medical Sciences, Beijing, China · Tianjin Key Laboratory of Ionic-Molecular Function of Cardiovascular Disease, Department of Cardiology, Tianjin Institute of Cardiology, The Second Hospital of Tianjin Medical University, Tianjin, China · State Key Laboratory of Vascular Homeostasis and Remodeling, NHC Key Laboratory of Cardiovascular Molecular Biology and Regulatory Peptides, Peking University, Beijing, China · Institute for Artificial Intelligence, Peking University, Beijing, China
Hyperkalemia is a life-threatening electrolyte disorder that is common in patients with chronic kidney disease and heart failure, yet frequent monitoring remains difficult outside hospital settings. We developed and validated Pocket-K, a single-lead AI-ECG system initialized from the ECGFounder foundation model for non-invasive hyperkalemia screening and handheld deployment. In this multicentre observational study using routinely collected clinical ECG and laboratory data, 34,439 patients contributed 62,290 ECG--potassium pairs. Lead I data were used to fine-tune the model. Data from Peking University People's Hospital were divided into development and temporal validation sets, and data from The Second Hospital of Tianjin Medical University served as an independent external validation set. Hyperkalemia was defined as venous serum potassium > 5.5 mmol/L. Pocket-K achieved AUROCs of 0.936 in internal testing, 0.858 in temporal validation, and 0.808 in external validation. For KDIGO-defined moderate-to-severe hyperkalemia (serum potassium >= 6.0 mmol/L), AUROCs increased to 0.940 and 0.861 in the temporal and external sets, respectively. External negative predictive value exceeded 99.3%. Model-predicted high risk below the hyperkalemia threshold was more common in patients with chronic kidney disease and heart failure. A handheld prototype enabled near-real-time inference, supporting future prospective evaluation in native handheld and wearable settings.
Figures & tables
Figure 1: Overview of Pocket-K development, validation, and proof-of-concept deployment. In the development stage, ECG– K+ pairs were constructed by linking each ECG to the nearest eligible venous serum potassium measurement within a ± 1-hour window. These paired data were used to fine-tune ECGFounder, a pretrained ECG foundation model, to develop Pocket-K. In the validation stage, model performance was assessed in three settings: an internal test set from the PKUPH development set, a temporal validation set at PKUPH, and an independent external validation set from SHTMU. In the proof-of-concept stage, a handheld device recorded a 30-s lead I ECG, which was processed through a smartphone-based workflow for near-real-time inference and generation of a hyperkalemia risk alert.
Characteristic
Development Set (PKUPH)
Temporal Validation Set (PKUPH)
Independent External Validation Set (SHTMU)
( N=10,409 ; n=26,145 )
( N=5,054 ; n=10,922 )
( N=18,976 ; n=25,223 )
Demographics
Age, years
54.08 (15.55)
54.60 (14.70)
58.25 (17.90)
Male sex, n (%)
5,642 (54.2%)
2,830 (56.0%)
9,994 (53.3%)
Laboratory measurements
Serum K+ , mmol/L
4.14 (0.36)
4.19 (0.40)
4.11 (0.42)
Table 1: Baseline clinical characteristics, ECG–potassium pairing quality, and technical specifications across study sets. Continuous variables are presented as mean (SD), and categorical variables as n (%). PKUPH: Peking University People’s Hospital; SHTMU: The Second Hospital of Tianjin Medical University; K+ : serum potassium; eGFR: estimated glomerular filtration rate; ECG: electrocardiogram.
Figure 2: ROC curves for model performance evaluation. (a) Internal testing ROC curve of PKUPH. (b) Temporal validation ROC curve of PKUPH. (c) Independent external ROC curve of SHTMU.
Figure 3: ROC curves for detection of KDIGO-defined moderate-to-severe hyperkalemia (serum potassium ≥ 6.0 mmol/L). (a) Temporal validation ROC curve in PKUPH. (b) Independent external validation ROC curve in SHTMU.
Figure 4: Signal-averaged waveform comparison. Average Lead I ECG waveforms are shown for patients correctly classified by the model as hyperkalemic (high risk, red line) and non-hyperkalemic (low risk, blue line). Shaded areas indicate standard deviation.
Figure 5: Clinical profiles of reference-negative samples stratified by model-predicted risk. In the external validation set, all samples shown had serum potassium values below the diagnostic threshold for hyperkalemia. They were stratified into a model low-risk group and a model high-risk group according to model output. Chronic kidney disease and heart failure were more common in the model high-risk group. The prevalence of chronic kidney disease increased from 1.2% in the model low-risk group to 3.2% in the model high-risk group ( P<0.001 ), and the prevalence of heart failure increased from 1.3% to 4.4% ( P<0.001 ).
Figure 6: Longitudinal tracking of serum potassium using single-lead AI-ECG. Representative patients showing the model’s ability to track within-patient potassium changes over time. The red line indicates laboratory serum potassium, and the blue dashed line indicates the model-predicted risk probability. Patient A shows progressive potassium rise with late threshold crossing. Patient B shows an early hyperkalemic episode followed by sustained recovery. Patient C shows recurrent severe fluctuation with corresponding changes in model output. Patient D shows gradual biochemical improvement accompanied by declining predicted risk. Overall, these examples suggest that model output tracked clinically meaningful potassium dynamics at the individual-patient level.
Figure 7: Proof-of-concept handheld deployment workflow of Pocket-K. (A) A handheld device paired with a smartphone application records a 30-s single-lead (Lead I) ECG. (B) The recorded signal is transmitted to the smartphone through Bluetooth and processed by a connected inference pipeline for near-real-time risk estimation. (C) Representative outputs from the prototype system show a low predicted hyperkalemia risk for a normokalemic example (serum potassium 4.1 mmol/L) and a high predicted risk for a severe hyperkalemia example (serum potassium 6.9 mmol/L).
Figure 8: STARD flowchart for study population selection and multicenter dataset partitioning. Panel A shows set construction at PKUPH, where 77,785 patients were initially screened. A total of 62,322 patients were excluded through a multistage quality-control pipeline: 29,522 had no ECG records, 32,542 had no eligible venous serum potassium measurement within a ± 1-hour window for ECG-anchored pairing, and 258 were removed because of poor data quality, including confirmed laboratory haemolysis or low signal-to-noise ratio. The remaining PKUPH population ( N=15,463 ; n=37,067 pairs) was partitioned chronologically at July 2021 into a development set ( N=10,409 ; n=26,145 pairs) and a temporal validation set ( N=5,054 ; n=10,922 pairs). Panel B shows independent external validation at SHTMU. Of 49,415 initially screened patients, 30,439 were excluded using the same criteria (missing ECGs: n=24,011 ; no eligible potassium measurement within ± 1 hour for ECG-anchored pairing: n=6,361 ; poor data quality: n=67 ), resulting in a final independent external validation set of 18,976 patients with 25,223 ECG– K+ pairs. In all panels, N denotes the number of unique patients and n denotes the number of ECG– K+ paired records.
Electrocardiogram classification uses machine learning to identify heart rhythms and cardiac conditions from ECG waveforms for screening, diagnosis, and monitoring.
Chagas disease screening via ECGs is limited by scarce and noisy labels in existing datasets. We propose a biomarker-based pretraining approach, where an ECG feature extractor is first trained to predict percentile-binned blood biomarkers from the MIMIC-IV-ECG dataset. The pretrained model is then fine-tuned on Brazilian datasets for Chagas detection. Our 5-model ensemble, developed by the Ahus AIM team, achieved a challenge score of 0.269 on the hidden test set, ranking 5th in Detection of Chagas Disease from the ECG: The George B. Moody PhysioNet Challenge 2025. Source code and the model are shared on GitHub: github.com/Ahus-AIM/physionet-challenge-2025
Elias Stenhede, Arian Ranjbar
Medical Technology & E-health, Akershus University Hospital, Lørenskog, Norway · Faculty of Medicine, University of Oslo, Oslo, Norway
This study presents AI-HEART, a cloud-based information system for managing and analysing long-duration ambulatory electrocardiogram (ECG) recordings and supporting clinician decision-making. The platform operationalises an end-to-end pipeline that ingests multi-day three-lead ECGs, normalises inputs, performs signal preprocessing, and applies dedicated deep neural networks for wave delineation, noise/quality detection, and beat- and rhythm-level multi-class arrhythmia classification. To address class imbalance and real-world signal variability, model development combines large clinically annotated datasets with expert-in-the-loop curation and generative augmentation for under-represented rhythms. Empirical evaluation on three-lead ambulatory ECG data shows that delineation accuracy is sufficient for automated interval measurement, noise detection reliably flags poor-quality segments, and arrhythmia classification achieves high specificity with clinically useful macro-averaged performance across common and rarer rhythms. Beyond predictive accuracy, AI-HEART provides a scalable deployment approach for integrating AI into routine ECG services, enabling traceable outputs, audit-friendly storage of recordings and derived annotations, and clinician review/editing that captures feedback for controlled model improvement. The findings demonstrate the technical feasibility and operational value of a noise-aware AI-ECG platform as a digital health information system.
KIOS Research and Innovation Center of Excellence (KIOS CoE), University of Cyprus, Panepistimiou 1, Aglantzia 2109, Cyprus · MEDTL Medical Technologies Ltd., 91 Limassol Avenue, Aglantzia, Office 202, Nicosia 2121, Cyprus · Department of Computer Science, University of Cyprus, Panepistimiou 1, Aglantzia 2109, Cyprus
Laboratory measurements are essential for clinical decision making but typically require invasive blood sampling. Recent research suggests that ECG recordings may contain information beyond cardiac rhythm and provide information related to laboratory values. This thesis investigates NT-proBNP prediction from ECG recordings and patient metadata using a multimodal deep learning framework, while considering the physiological plausibility of ECG-based prediction. It also compares a 1-lead configuration using lead II with the corresponding 12-lead configuration. Additional experiments on glucose, hemoglobin, and potassium are presented in the appendix. Regression, binary classification, and three-class classification experiments were conducted using MIMIC-IV, with the three-class task based on clinically established thresholds. The 12-lead configuration consistently achieved better performance. Regression reached a mean absolute error of 534.98 pg/mL for 1-lead and 514.27 pg/mL for 12-lead. Binary classification achieved AUROCs of 0.849 and 0.876, respectively, while three-class classification achieved macro-averaged AUROCs of 0.738 and 0.752. The results indicate that precise NT-proBNP regression remains challenging, whereas broader concentration ranges can be classified more effectively. The findings are consistent with the physiological hypothesis that ECG characteristics and NT-proBNP concentrations reflect shared cardiovascular processes. ECG-based NT-proBNP prediction may therefore provide complementary information, particularly for classifying broader concentration ranges. However, laboratory measurements remain necessary, and validation on independent datasets is required.
Walentin Palanchuk
Carl von Ossietzky Universität Oldenburg Fach-Bachelor Informatik