Continuous Determination of Respiratory Rate in Hospitalized Patients using Machine Learning Applied to Electrocardiogram Telemetry
Thomas Kite PhD, Brian Ayers MD MBA, Nicholas Houstis MD PhD, Asishana A. Osho MD, Thoralf M. Sundt MD, Aaron D Aguirre MD PhD · 2025
Read Continuous Determination of Respiratory Rate in Hospitalized Patients using Machine Learning Applied to Electrocardiogram Telemetry by Thomas Kite PhD, Brian Ayers MD MBA, Nicholas Houstis MD PhD, Asishana A. Osho MD, Thoralf M. Sundt MD, Aaron D Aguirre MD PhD (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-15947.txt.
Introduction
Thomas Kite and five co-authors trained a neural network to read breathing rate from heart telemetry in hospitalized patients in a 2025 paper. Breathing rate is a key vital sign, yet counts by hand are slow and often wrong. Steady monitoring exists in intensive care but not on most wards.
Heart traces carry several faint traces of breathing, which the network learns to read. Its error stayed under two breaths per minute across test sets from the same and other sources. In past cases that ended in breathing failure, the steady trace tracked the slide toward a breathing tube. The authors see this as a step toward a hospital-wide early warning system built on gear already in place.
Catalog record
| CREATOR | Thomas Kite PhD, Brian Ayers MD MBA, Nicholas Houstis MD PhD, Asishana A. Osho MD, Thoralf M. Sundt MD, Aaron D Aguirre MD PhD |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 35 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-15947
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Continuous Determination of Respiratory Rate in Hospitalized Patients using Machine Learning Applied to Electrocardiogram Telemetry Thomas Kite PhD, Brian Ayers MD MBA, Nicholas Houstis MD PhD, Asishana A. Osho MD, Thoralf M. Sundt MD, Aaron D Aguirre MD PhD Abstract Respiration rate (RR) is an important vital sign for clinical monitoring of hospitalized patients, with changes in RR being strongly tied to changes in clinical status leading to adverse events. Human labels for RR, based on counting breaths, are known to be inaccurate and time consuming for medical staff. Automated monitoring of RR is in place for some patients, typically those in intensive care units (ICUs), but is absent for the majority of inpatients on standard medical wards who are still at risk for clinical deterioration. This work trains a neural network (NN) to label RR from electrocardiogram (ECG) telemetry waveforms, which like many biosignals, carry multiple signs of respiratory variation. The NN shows high accuracy on multiple validation sets (internal and external, same and different sources of RR labels), with mean absolute errors less than 1.78 breaths per minute (bpm) in the worst case. The clinical utility of such a technology is exemplified by performing a retrospective analysis of two patient cohorts that suffered adverse events including respiratory failure, showing that continuous RR monitoring could reveal dynamics that strongly tracked with intubation events. This work exemplifies the method of combining pre-existing telemetry monitoring systems and artificial intelligence (AI) to provide accurate, automated and scalable patient monitoring, all of which builds towards an AI-based hospital-wide early warning system (EWS). keywords Machine Learning, Clinical Monitoring, Electrocardiogram Figure 1: Graphical abstract. A) A retrospective archive of standard vital signs and continuous telemetry waveforms is used for analysis. B) Downstream training and validation datasets are generated by aligning ECG telemetry waveforms with labels consisting of minute-long averages of automated respiration rate sources. C) Final datasets included a typical train/tune/test split from the same hospital and same respiration source, stratified by patient. Further validations include a distinct respiration source for ground truth respiratory rate from the same institution and a separate cohort from an external institution. D) A neural network is trained to predict the average respiration rate from the ECG telemetry segments, with training by backpropagating the mean square error. E) The neural network is technically validated by inspecting labels and predictions for all test sets. F) The neural network is clinically validated by studying two clinical cohorts: a cohort of patients that experience clinical deterioration requiring a rapid response event, and a cohort of patients who required a second intubation following thoracic surgery. 1 Introduction The modern hospital setting is a highly collaborative and data rich environment which aims to closely monitor and promote the recovery of each and every hospitalized patient. Despite the sophistication of these environments, a persistently high number of patients suffer unexpected deterioration while under observation, leading to deaths that are potentially preventable [ 1 , 2 ] . It is tempting to tackle this problem through more intensive monitoring with a broader range of tests or with a higher frequency of assessments, but resource constraints already spread medical staff thin, increasing the possibility of error and raising the risk of burnout [ 3 ] . This work proposes an automated, inexpensive and scalable method for monitoring a key vital sign – respiration rate [ 4 ] (RR) – on all hospitalized patients with access to standard electrocardiogram (ECG) telemetry. Such an approach would not only avoid burdening medical staff with new responsibilities, but could actually relieve them of manually recording RR, a measurement that is both time consuming and prone to inaccuracy [ 5 , 6 ] . Respiratory rate is a key determinant of blood oxygen delivery and carbon dioxide removal, one whose autoregulation can reveal important information about insults to a patient’s respiratory and metabolic homeostasis. For example, a patient with hypoxic respiratory failure will breathe faster in an effort to defend blood oxygen levels. Similarly, a patient with a primary metabolic acidosis, perhaps from poor perfusion and rising lactic acid levels in the blood, will breathe faster to expel blood carbon dioxide and thereby compensate for this acidosis. Moreover, loss of respiratory control, as can result from neurologic injury or from medications that depress respiration such as opiates, can lead to inappropriately low RR and risk for hypoventilation. A single metric, RR, thus provides a window into a spectrum of physiologic processes, making it an essential vital sign to monitor patients for deterioration. For these reasons and more, RR is a commonly used predictor in hospital early warning systems (EWS) [ 7 , 8 ] , defined here as collections of monitoring systems and algorithms to analyze patient data and predict deterioration as early as possible. Devices for the automated monitoring of RR are available, but their use is often restricted to patients in the intensive care unit (ICU), limiting their applicability to a general hospital-wide EWS. Previous studies have demonstrated the potential to quantify respiratory rate from various biosignals (most frequently photoplethysmography, PPG, but also ECG) [ 9 ] . In the hospital setting, ECG telemetry is a ubiquitous technology that allows for the continuous, noninvasive monitoring of each patient’s cardiac electrical activity. Automated systems already use the ECG to track heart rate and rhythm, which allows for some simple alarms. However, telemetry monitoring systems also capture the full high-resolution ECG waveform, beyond simple timing intervals, making it a promising target for the application of artificial intelligence (AI) algorithms to augment patient care [ 10 ] . Importantly for this work, the ECG telemetry signal contains variations attributable to patient respiration [ 11 ] , which can be leveraged using a neural network (NN) trained to predict RR. Such an AI model could be connected to existing hospital telemetry systems, enabling real-time RR monitoring for almost all hospitalized patients. In this work we train a model to predict RR from ECG telemetry by leveraging the two largest telemetry databases used in scientific work to date (to the best of the authors’ knowledge). Technical and clinical validation establishes the model as both an accurate predictor of RR and a useful metric of clinical deterioration (Fig. 1 ). The results advocate for a future of AI-enhanced monitoring systems that could enable accurate and low-burden EWS’s that improve patient outcomes. 2 Dataset & cohort characteristics This work uses a retrospective archive of telemetry biosignals and vital signs collected from Massachusets General Hospital (MGH) between 2014 and 2023 (study approved under Institutional Review Board protocol 2020P003053). Respiratory rate measurements from either impedance pneumography (ImP) or capnography are time-aligned with single-lead ECG telemetry and divided into 60-second segments (a single lead is used as a single model input, taken from from any available physical leads, usually I, II, III and V). A comparable dataset was also curated from the open source MIMIC III waveform database [ 12 ] as an external validation. More details of the selection criteria and curation are provided in Sect. 5.1 . Key patient characteristics of the datasets and their partitions are given in Table 1 . The first three columns report patient-level information such as unique patient count, age distribution and patient sex.