AI-based detection of worsening heart failure from low-resolution telemonitoring data

Erik Aerts; Yinan Yu; Annika Rosengren; Michael Fu; Martin Lindgren; Falk Dippel; Martin Adiels; Helen Sjöland

cs.AISep 24, 2026

AI-based detection of worsening heart failure from low-resolution telemonitoring data

Authors: Erik Aerts, Yinan Yu, Annika Rosengren, Michael Fu, Martin Lindgren, Falk Dippel, Martin Adiels, Helen Sjöland

Organizations: Department of Computer Science and Engineering, Chalmers Universiy of Technology, Gothenburg, Sweden · Department of Computer Science and Engineering, University of Gothenburg, Gothenburg, Sweden · Center for Digital Health, Sahlgrenska University Hospital, Gothenburg, Sweden · Departement of Molecular and Clinical Medicine, Sahlgrenska Academy, University of Gothenburg, Gothenburg, Sweden · Department of Medicine, Geriatrics and Emergency Medicine, Sahlgrenska University Hospital, Gothenburg, Sweden · Institute of Medicine, Sahlgrenska Academy, University of Gothenburg, School of Public Health and Community Medicine, Gothenburg, Sweden

Abstract

Objective: Heart failure (HF) presents a healthcare challenge due to its high comorbidity burden, aging patient population and frequent hospitalizations. Remote monitoring offers a promising approach to managing HF patients by early detection of health deterioration. Developing autonomous systems to detect signs of worsening in telemonitoring data is of interest to reduce the workload of healthcare personnel. Methods: We propose the TRACER model, a Transformer with Contrastive Event Representation, designed to predict timelines leading to rare hospitalization events in low-resolution and irregularly sampled telemonitoring data. TRACER incorporates time-aware embeddings for each biomarker, contrastive pre-training to enhance anomaly detection via representation learning, and independent binary classifiers for detection. We used measurement data containing remotely recorded biomarker sequences from 276 HF patients segmented into overlapping windows based on temporal rules, and labeled the windows based on the occurrence of HF relevant hospitalizations at the latter edge of the window. Results: TRACER was able to correctly predict 66.7% timelines leading up to HF hospitalizations in the highly imbalanced real-world dataset with an overestimation of 7.9%. Reformulating the training of TRACER as an event detection problem improved the predictive performance compared with training directly on forecasting windows, enabling more effective use of the limited hospitalization events. Conclusion: TRACER demonstrated superior performance in detecting signs of worsening status in real-world telemonitoring data compared to the other tested models. Significance: TRACER shows promise in identifying signs of clinical deterioration that allow for alerts to be generated to provide counteractive treatment in patients with HF.

Figures & tables

Fig. 1 : Data overview for monitored biomarkers in the telemonitoring. From the left: (a) mean ( $\mu$ ) and standard deviation ( $\sigma^{2}$ ) of measurement counts per patient; (b) the average number of days between measurements; (c) the average number of days in the monitoring period per patient

Fig. 2 : Illustration of the window process. Four windows from the recorded variables measurement signal for a patient (top) are highlighted with bars, and the location of the patient during the period is showcased with black (home) and red (hospital). Windows 1 and 2 show windows created from an overlapping region of measurements. One out of four windows does not contain sufficient data points and is therefore discarded (window 3). One out of four windows contains a hospital event (red) at the latter edge and is therefore labeled with label 1 accordingly (window 4).

Fig. 3 : Illustration of telemonitoring pipeline. The patient at home measures daily recordings of their biomarkers with the provided devices (top left). The data is joined with previous recordings and passed, alongside a representation of the patient profile, to the DL model hosted on a cloud-based service (bottom). The output of the model is sent to the hospital (top right) to provide an indication of the health status of the patient to clinicians.

Fig. 4 : Illustration of the model architecture. The model is divided into three parts: transformer encoders (left), pre-training (middle) and classification networks (right). The embedding process is shown in the red-striped box. The tokenized measurement sequence is shown as boxes labeled $M_{i}$ and embeddings are shown as boxes labeled $E_{i}$ . Each biomarker is illustrated as a different geometrical shape as the output from the encoders. For ABC pre-training (top), each representation is projected into a specific feature domain where the hospitalized windows (red) are separated from the non-hospitalized windows (white). For PBC pre-training (bottom), all representations are projected into the same domain where each window from the same patient is regarded as a positive pair (green link) and each window from different patients is seen as a negative pair (red link). Abbreviations: t = Time, HR = Heart Rate, SYS = Systolic Blood Pressure, DIA = Diastolic Blood Pressure, M = Measurement

Fig. 5 : Illustration of the embedding process. The sources of information are shown to the left. Measurement

M_{1}

in the figure is decomposed into a value and a timestamp. The value undergoes affine transformation to produce a value embedding, whereas the timestamp is converted into a time embedding with sinusoidal encoding. These are then further transformed to form position-aware embeddings followed by concatenation. The patient profile is independently embedded through an affine layer. All components: a CLS -token, the patient profile, and the measurement embeddings are concatenated to form the input sequence. Abbreviations: t = time, M = measurement

Number of windows	51030
Number of non-hospital windows	50989
Number of hospital windows	41
Average window length (measurements)	13.58
Average delay between measurements (days)	0.957

TABLE I : Overview of the properties of the constructed time windows of measurements in both datasets. All values presented in the table are calculated without up- or downsampling.

Model type (Pre-training)	Parameter count	Sensitivity	Specificity	FA-PPM	NNE
Transformer (No pre-train)	25837412	0.629	0.910	2.86	24.87
TRACER	25837412	0.667	0.921	2.49	20.79
Transformer (PBC)	25837412	0.037	0.937	2.01	286.53
Transformer (PBC&ABC)	25837412	0.518	0.909	2.86	30.14
Transformer (ABC&PBC)	25837412	0.370	0.935	2.04	30.00
LSTM	253635	0.480	0.932	2.14	16.12

TABLE II : Overview of the performance of predicting timelines leading to hospitalization for the tested architectural setups trained on event-based detection in the study. Abbreviations: PBC = Patient-based contrastive, ABC = Anomaly-based contrastive, ARIMA = Autoregressive integrated moving average.

Model type (Pre-training)	Sensitivity	Specificity	FA-PPM	NNE
Transformer (No pre-train)	0.222	0.905	2.99	71.94
TRACER	0.259	0.976	0.76	16.58
Transformer (PBC)	0.037	0.978	0.69	98.03
Transformer (PBC&ABC)	0.111	0.945	1.73	83.33
Transformer (ABC&PBC)	0	0.985	0.47	-
LSTM	0.149	0.945	1.73	79.44

TABLE III : Overview of the performance of predicting timelines leading to hospitalization for the tested architectural setups trained on timeline prediction in the study. Abbreviations: PBC = Patient-based contrastive, ABC = Anomaly-based contrastive, ARIMA = Autoregressive integrated moving average

	Biomarker	Sensitivity	Specificity	FA-PPM	NNE
One at	HR	0.667	0.924	2.39	20.16
a time	SYS	0.667	0.917	2.61	21.69
	DIA	0.667	0.921	2.48	20.70
	W	0.111	0.961	1.23	56.18
	Model
Combination	TRACER + W	0.389	0.927	2.29	30.39

TABLE IV : Overview of the performance of predicting timelines leading to hospitalization for each biomarker and with added weight measurements in the study with TRACER. Abbreviations: HR = Heart Rate, SYS = Systolic Blood Pressure, DIA = Diastolic Blood Pressure, W = Weight

Patient #	Sensitivity (TRACER)			Sensitivity (Transformer)
Biomarker	HR	SYS	DIA	HR	SYS	DIA
Patient 1	0	0	0	0	0	0
Patient 2	0.833	0.833	0.833	0.833	0.833	0.667
Patient 3	1	1	1	1	1	1

TABLE V : Overview of the biomarker-specific sensitivity of predicting timelines leading to hospitalization for each hospitalized patient in the validation set. Abbreviations: HR = heart Rate, SYS = systolic blood pressure, DIA = diastolic blood pressure

# Blocks	Parameters	Sensitivity	Specificity	FA-PPM	NNE
1	8976740	0.185	0.886	3.59	103.30
2	17407076	0.481	0.924	2.39	26.88
3	25837412	0.667	0.921	2.49	20.79
4	34267748	0.704	0.919	2.55	20.20
5	42698084	0.667	0.918	2.58	21.46

TABLE VI : Overview of the performance of predicting timelines leading to hospitalization for different numbers of attention blocks in the transformer encoders.

Metric (mean $\pm$ std)	TRACER	Transformer (No pre-train)
Sensitivity	0.537 $\pm$ 0.15	0.441 $\pm$ 0.13
Specificity	0.921 $\pm$ 0.04	0.931 $\pm$ 0.03
FA-PPM	2.50 $\pm$ 1.12	2.15 $\pm$ 1.07
NNE	101.10 $\pm$ 67.66	123.09 $\pm$ 110.20

TABLE VII : Overview of the average performance for five different seeds. Abbreviations: std = standard deviation

Explore similar work

Narrative Feature or Structured Feature? A Study of Large Language Models to Identify Cancer Patients at Risk of Heart Failure
Mar 18, 2024Ziyi Chen, Mengyuan Zhang, Mustafa Mohammed Ahmed +4Cardiac Electronic Health Records
Department of Health Outcomes and Biomedical Informatics, College of Medicine, University of Florida, Gainesville, FL, USA · Division of Cardiovascular Medicine, Department of Medicine, College of Medicine, University of Florida, Gainesville, FL, USA · Division of Hematology & Oncology, Department of Medicine, College of Medicine, University of Florida, Gainesville, FL, USA
Artificial Intelligence-Assistant Cardiotocography: Unified Model for Signal Reconstruction, Fetal Heart Rate Analysis, and Variability Assessment
May 14, 2026Xiaohua Wang, Kai Yu, XuXiao Liang +2Signal Reconstruction
Artificial Intelligence Research Center, Bengbu Medical University, Bengbu 233000, China. · CHARMMIRAEL Biotech Co., Ltd, Nanjing 210000, China.
DT-Transformer: A Foundation Model for Disease Trajectory Prediction on a Real-world Health System
May 14, 2026Yunying Zhu, Andrew R Weckstein, Kueiyu Joshua Lin +1Readmission Prediction Electronic Health Records
Division of Pharmacoepidemiology and Pharmacoeconomics, Department of Medicine, Brigham and Women’s Hospital, Harvard Medical School, Boston, MA, USA · Harvard T.H. Chan School of Public Health, Harvard University, Boston, MA, USA · Kempner Institute for the Study of Natural and Artificial Intelligence, Harvard University, MA, USA +2

Mar 18, 2024, cs.LGJ/K move · Enter open · S save

Narrative Feature or Structured Feature? A Study of Large Language Models to Identify Cancer Patients at Risk of Heart Failure

Ziyi Chen, Mengyuan Zhang, Mustafa Mohammed Ahmed, Yi Guo+3

Department of Health Outcomes and Biomedical Informatics, College of Medicine, University of Florida, Gainesville, FL, USA · Division of Cardiovascular Medicine, Department of Medicine, College of Medicine, University of Florida, Gainesville, FL, USA · Division of Hematology & Oncology, Department of Medicine, College of Medicine, University of Florida, Gainesville, FL, USA

Cardiac Electronic Health Records Multidisciplinary Tumor Boards Feature Engineering Cnn-Lstm Narratives

Open paper AskPDF arXiv

Cancer treatments are known to introduce cardiotoxicity, negatively impacting outcomes and survivorship. Identifying cancer patients at risk of heart failure (HF) is critical to improving cancer treatment outcomes and safety. This study examined machine learning (ML) models to identify cancer patients at risk of HF using electronic health records (EHRs), including traditional ML, Time-Aware long short-term memory (T-LSTM), and large language models (LLMs) using novel narrative features derived from the structured medical codes. We identified a cancer cohort of 12,806 patients from the University of Florida Health, diagnosed with lung, breast, and colorectal cancers, among which 1,602 individuals developed HF after cancer. The LLM, GatorTron-3.9B, achieved the best F1 scores, outperforming the traditional support vector machines by 39%, the T-LSTM deep learning model by 7%, and a widely used transformer model, BERT, by 5.6%. The analysis shows that the proposed narrative features remarkably increased feature density and improved performance.

Abstract

Explore similar work

Narrative Feature or Structured Feature? A Study of Large Language Models to Identify Cancer Patients at Risk of Heart Failure

Artificial Intelligence-Assistant Cardiotocography: Unified Model for Signal Reconstruction, Fetal Heart Rate Analysis, and Variability Assessment

DT-Transformer: A Foundation Model for Disease Trajectory Prediction on a Real-world Health System