PaperlayerView exact revision on arXiv

Source-linked AI summary

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

arXiv:2609.29742v1cs.AI

TL;DR

TRACER targets early identification of worsening heart failure from low-resolution, irregular telemonitoring data. It predicted 66.7% of timelines leading to hospitalization with 7.9% overestimation, and event-based training made more effective use of limited hospitalization events.

  • Problem

    Heart failure’s frequent hospitalizations, comorbidities, and aging patient population motivate telemonitoring systems that can detect worsening health status.

    Source: §I · p. 1

  • Method

    TRACER uses time-aware transformer encoders, contrastive pre-training, and biomarker-specific binary classifiers to predict timelines leading to hospitalization from telemonitoring data.

    Source: §I · p. 2

  • Results

    66.7% of timelines leading to hospitalization events were correctly predicted, with a 7.9% overestimation rate.

    Source: §I · p. 2

  • Takeaways & Limitations

    Event-based training improved identification of hospitalization timelines and enabled more effective learning from sparse telemonitoring data.

    Source: §VI · p. 10, §VII · p. 11

    TRACER is intended for risk stratification, not acute-deterioration diagnosis, and hospitalization patterns may combine physiological and other temporal changes.

    Source: §VI · p. 10

Abstract

from arXiv · show

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.

I. INTRODUCTION

Heart failure creates substantial healthcare burden, while telemonitoring systems often lack temporal context for detecting worsening. The paper introduces TRACER and reports its performance on predicting timelines before HF-related hospitalization.

I. INTRODUCTION

  • Heart failure brings frequent hospitalizations alongside comorbidities and older age, creating a significant healthcare burden.The introduction notes these challenges are expected to increase with population aging and accumulating long-term conditions.Source: §I · p. 1
  • Many telemonitoring systems rely on recent recordings or threshold alerts, often ignoring temporal dependencies and trends.The paper notes that detecting worsening and alerting clinicians could reduce healthcare personnel workload.Source: §I · p. 1
  • TRACER combines transformer encoders, time-aware biomarker embeddings, contrastive pre-training, and binary classifiers to predict hospitalization timelines.Reformulating training as event detection can make more effective use of limited hospitalization events in an imbalanced dataset.Source: §I · p. 2
  • 66.7% of timelines leading to hospitalization events were correctly predicted, with a 7.9% overestimation rate.The reported results use a highly imbalanced real-world telemonitoring dataset.Source: §I · p. 2

A. Telemonitoring

The paper situates TRACER among remote heart-failure monitoring approaches, irregular medical time-series methods, and contrastive-learning techniques. It targets low-resolution, irregular measurements and rare-event prediction.

A. Telemonitoring

  • Implantable sensors provide richer physiological information but apply to a selected HF population, while non-invasive monitoring and routine measurements offer broader reach with deployment and workflow challenges.The paper notes that dedicated devices and infrastructure may limit scalability, and staff may need to review large volumes of incoming data.Source: A. Telemonitoring · p. 2

B. Medical time series

  • Irregular medical time series can treat sampling gaps as missing data or model elapsed time, but their use for rare events in low-resolution telemonitoring remains less explored.Prior approaches include time-aware recurrent models and methods encoding temporal and spectral representations.Source: B. Medical time series · pp. 2–3

C. Contrastive learning

  • Contrastive learning organizes similar and dissimilar data pairs, and TRACER applies contrastive objectives to learn patterns from irregular, low-resolution telemonitoring data under high class imbalance.Prior methods include contrastive frameworks for images, multimodal data, medical time series, and anomaly detection.Source: C. Contrastive learning · p. 3

III. TELEMONITORING DATA

The study retrospectively analyzes six months of telemonitoring and hospitalization records from patients with heart failure. It forms overlapping measurement windows and labels those preceding HF-relevant hospitalization events.

III. TELEMONITORING DATA

  • Patients were monitored for six months across three hospitals, with daily recordings of heart rate and blood pressure; inconsistently recorded body weights were excluded from model development.The analysis was retrospective, and model outputs were not available to clinicians during monitoring or used to trigger care actions.Source: §III · p. 4
  • Windows contained up to 14 measurements from a 21-day range, were discarded if fewer than 10 points were available, and received a positive label for HF-related hospitalization within 24 hours after the last measurement.Windows overlapped, and both measurement values and timestamps were normalized before padding sequences to length 14.Source: §III · p. 4
  • The envisioned model output would be reviewed by clinicians or nurses to prompt possible deterioration assessment or intensified follow-up.In this study, model outputs were not provided to the clinical team during monitoring.Source: §III · p. 4

IV. METHODS

TRACER uses biomarker-specific transformer encoders and contrastive pre-training before binary classification. The methods also describe its input windows and patient-based contrastive pre-training.

A. Deep learning model

  • The data overview reports measurement-count mean and standard deviation, average days between measurements, and average monitoring duration; overlapping windows with insufficient data are discarded.A window containing a hospital event at its latter edge is labeled positive.Source: Figure 1 · p. 5, Figure 2 · p. 5
  • Three biomarker-specific transformer encoders use self-attention, and two-layer classifiers produce binary hospitalization predictions from enriched embeddings.Each encoder contains three attention blocks; the model architecture also allows patient-based and anomaly-based contrastive pre-training.Source: A. Deep learning model · pp. 4–5

B. Contrastive pre-training

  • Patient-based contrastive pre-training maps embeddings through two-layer projection heads and compares similarity in a structured space.Similarity matrices are padded for varying biomarker batch sizes, and a sharpened sigmoid modulates the scores.Source: 1) Patient-based contrastive · p. 5

1) Patient-based contrastive:

Patient-based contrastive pre-training uses patient-profile similarity to define positive window pairs, then shapes the window representations with a pairwise loss.

1) Patient-based contrastive:

  • Windows whose patient profiles meet the ≥99% similarity threshold are treated as positive pairs.The threshold is θ, and profile similarity can apply across windows from the same or different biomarkers.Source: 1) Patient-based contrastive · p. 6
  • The PBC loss is defined over available lower-dimensional window representations, with a trainable temperature sharpening the sigmoid distance calculation.The indicator marks a positive pair when profile-vector cosine similarity reaches θ.Source: 1) Patient-based contrastive · p. 6

2) Anomaly-based contrastive:

Anomaly-based contrastive pre-training uses hospitalization-window labels to define pairings and compute a loss over projected window representations.

2) Anomaly-based contrastive:

  • The ABC loss is defined over available lower-dimensional window representations using a sharpened sigmoid function.Its cosine-similarity scaling parameter is γ, and τa is a trainable temperature parameter.Source: 2) Anomaly-based contrastive · p. 6
  • ABC pairings are positive when window labels match and negative when they differ.The indicator assigns yij = 1 for equal labels and 0 otherwise.Source: 2) Anomaly-based contrastive · p. 6

C. Embedding

The embedding design combines measurements’ values, timestamps, and sequence positions with a separately embedded patient profile. Comparison models include ARIMA/XGBoost and two LSTM variants.

C. Embedding

  • Measurements receive value, time, and positional embeddings before concatenation; the patient profile is projected separately, and a random CLS-token starts the input sequence.Values are linearly projected; timestamps use sinusoidal encoding of relative UNIX time, with absolute positional embeddings added to value and time embeddings.Source: C. Embedding · pp. 6–7, Figure 5 · p. 7

D. Comparison models

  • ARIMA/XGBoost uses window-specific ARIMA parameters alongside measurement values, excluding windows when ARIMA fails to produce valid parameters.The ARIMA models use previous, current, and next windows; XGBoost is configured with up to 200 trees and depth 7.Source: D. Comparison models · p. 7
  • The comparison includes a four-layer bidirectional LSTM and a T-LSTM that models elapsed time with a decay mechanism on the cell-state short-term component.The first LSTM processes each biomarker separately; T-LSTM uses measurements and time gaps, then classifies from its final hidden state.Source: D. Comparison models · p. 7

V. EXPERIMENTS

The experiments use a patient-independent training and validation split on a highly imbalanced cohort, with window construction and alert-burden measures specified.

A. Patient overview

  • 20 HF-relevant admissions from 18 individuals were recorded among 103 total admissions from 56 individuals.The cohort comprised 276 patients; the class imbalance ratio was 1244.63 before training-set resampling.Source: A. Patient overview · p. 7

B. Experimental setup

  • The 90/10 training-validation split was patient-independent, and the training set was resampled while validation retained the original class distribution.Windows had 75% overlap; duplicate hospitalization windows were added and 20% of non-hospitalization windows discarded in training.Source: B. Experimental setup · pp. 7–8
  • Sensitivity and specificity assess timeline prediction, while FA-PPM and NNE quantify false-alert and alert-evaluation burden.FA-PPM divides false-positive alerts by cohort patient-months; NNE divides positive alerts by correctly detected hospitalization events.Source: B. Experimental setup · p. 8

C. Main results

The results compare event-based detection with direct timeline prediction and assess how contrastive pre-training affects TRACER. Event-based training exposed models to multiple views of rare hospitalization events, while direct timeline prediction produced lower sensitivity.

1) Event-based training for timeline prediction:

  • The un-pretrained transformer had higher sensitivity than LSTM and ARIMA/XGBoost models with similar overestimation, while all models generated about two to three false alerts per patient-month.Across models, NNE ranged from 15.9 for T-LSTM to 286.5 for the PBC pre-trained transformer.Source: 1) Event-based training for timeline… · p. 8, Table II · p. 9

C. Main results

  • Event-based training allowed hospitalization events to occur anywhere within overlapping windows, giving models multiple views of the few training events.This formulation aimed to enrich the models’ understanding of contextual relationships around events.Source: 1) Event-based training for timeline… · p. 8

1) Event-based training for timeline prediction:

  • ABC pre-training increased timeline prediction by 3.8 percentage units and reduced hospitalization overestimation by 1.1 percentage units versus no pre-training.PBC pre-training substantially decreased performance; both combined pre-training orders had lower sensitivity than TRACER.Source: 1) Event-based training for timeline… · p. 8

2) Training with timeline prediction:

  • Direct timeline-prediction training reduced sensitivity across architectures versus event-based training; TRACER’s lower overestimation, FA-PPM, and NNE came with a 61.1% sensitivity reduction.The ABC&PBC pre-trained transformer detected no validation timeline leading to hospitalization.Source: 2) Training with timeline prediction · p. 8, Table III · p. 9

D. Ablation studies

The ablations assess the contribution of individual biomarkers, patient-level detection, and encoder depth. Cardiovascular measurements performed similarly, while weight-only and weight-added configurations were less favorable.

1) Performance with added weight measurements:

  • Adding weight to HR, SYS, and DIA reduced correct timeline identification and increased NNE, with only a minor decrease in overestimation.Weight-only TRACER did not detect hospitalization events on par with models using cardiovascular physiology variables.Source: 1) Performance with added weight… · p. 9, Table IV · p. 9

2) Patient-based hospitalization detection:

  • Among four validation patients, one lacked sufficient pre-hospitalization measurements; both models detected Patients 2 and 3, while Patient 1 was missed across all biomarkers.ABC pre-training increased Patient 2’s DIA sensitivity, contributing to the overall sensitivity increase.Source: 2) Patient-based hospitalization… · p. 9, Table V · p. 9
  • One attention block gave the lowest performance; three or more blocks performed similarly, with four marginally better at detecting hospitalization timelines.The study varied the number of attention blocks in each encoder.Source: 3) Performance with varying encoder… · p. 9

3) Performance with varying encoder complexity:

TRACER’s performance benefited from event-based training, while results also depended on encoder complexity and the pre-training strategy. Interpretation is constrained by alert burden, cohort scope, and uncertain generalizability.

4) Performance across varying seeds:

VI. DISCUSSION

  • Event-based training improved TRACER’s identification of hospitalization timelines; TRACER also outperformed the tested transformer and LSTM models.Using multiple temporal contexts around limited hospitalization events may support learning from sparse telemonitoring data.Source: §VI · p. 10

3) Performance with varying encoder complexity:

  • Performance improved with additional encoder blocks up to three; one-block TRACER performed worse than the ARIMA/XGBoost baseline.Additional blocks beyond three provided limited improvement; the model’s approximately 10^7 parameters also affect training and runtime usability relative to an LSTM with roughly 250,000 parameters.Source: §VI · p. 10, Table VI · p. 10

VI. DISCUSSION

  • Weight performed poorest alone and reduced performance when added to the three cardiovascular variables, but inconsistent recording and few events qualify that result.Heart rate and blood pressure contributed more strongly in this cohort, which largely comprised patients with newly diagnosed HF undergoing therapy optimization.Source: §VI · p. 10
  • With an imbalance ratio of 1244.63, TRACER generated about 2.5 false alerts per patient-month and required about 21 clinical evaluations per true hospitalization.The authors note possible cognitive overload and alert fatigue, while comparing this burden with routine review of every daily measurement by an HF nurse.Source: §VI · p. 10
  • The authors frame TRACER as risk stratification, not acute-deterioration diagnosis; only 18 of 276 patients had HF-related hospitalization during six months.The cohort was comparatively stable, and the authors emphasize accessible longitudinal monitoring without additional procedures or implantable infrastructure.Source: §VI · p. 10
  • ABC pre-training enhanced prediction, whereas PBC generally reduced it; applying PBC before ABC yielded worse sensitivity than LSTMs.The authors suggest that few subjects and events, plus a 99% similarity threshold applied to padded profiles, may produce misleading positive pairs.Source: §VI · p. 11
  • Generalizability to other healthcare settings, populations, and devices remains uncertain; the authors call for external validation, local calibration, and prospective evaluation.Differences in case mix, workflows, measurement routines, device characteristics, and data completeness may affect performance.Source: §VI · p. 11

VII. CONCLUSION

TRACER outperformed the other tested ML- and DL-based implementations for predicting timelines leading to worsening HF in low-resolution telemonitoring data. Pre-training could improve or harm performance, while event-detection training made more effective use of limited hospitalization events.

VII. CONCLUSION

  • TRACER demonstrated superior performance to the other tested ML- and DL-based implementations for predicting timelines leading to worsening HF in low-resolution telemonitoring data.The conclusion reports that pre-training both boosted and harmed performance, while reformulated event-detection training made more effective use of limited hospitalization events in the imbalanced real-world dataset.Source: §VII · p. 11
  • Pre-training techniques were observed to both boost and harm TRACER’s performance.Source: §VII · p. 11
  • Training TRACER on a reformulated event-detection task enabled more effective use of limited hospitalization events in the imbalanced real-world dataset.Source: §VII · p. 11

Related papers

Loading 2609.29742v1…