Machine learning outperforms state-of-the-art continuous vital sign monitoring

Machine learning outperforms state-of-the-art continuous vital sign monitoring | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Machine learning outperforms state-of-the-art continuous vital sign monitoring Norman K. Pedersen, Andreas Hasselriis, Carl I. Askehave, Jesper Mølgaard, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8436885/v1 This work is licensed under a CC BY 4.0 License Status: Under Revision Version 1 posted 14 You are reading this latest preprint version Abstract Continuous Vital Sign Monitoring (CVSM) may allow early detection of patient deterioration with clinical impact from reduced complications. CVSM-based alert algorithms are vulnerable to missing data and artefacts, where large numbers of false positive alerts impose alert fatigue among healthcare staff, especially in low-staffed environments such as general wards. An unexplored option for overcoming missing data and improving alert precision may be using machine learning (ML) models with linear regression to summarize periods of patient data to capture critical vital sign trajectories. Four different ML-analyses of 5 unique vital signs were compared to a state-of-the-art algorithm combining vital sign duration and severity. The data consisted of continuous and semi-continuous vital signs alongside timestamps for physician-curated Serious Adverse Events (SAE) and patient metadata, from 2423 patients monitored during admission for major surgery or acute medical disease. Our results demonstrate that ML-based models improve True Positive Rates (TPR by 0.4) and False Positive Rates (FPR by 0.69). Compared to threshold-based alerts, our approach, showed superior performance in predicting SAE within 24 hours (p < 0.001) and 8 hours (p < 0.001) of an alert. The substantial improvements validate our models’ competitive edge on clinical data, when directly compared to current alerting systems, without increasing false positives. Biological sciences/Computational biology and bioinformatics Health sciences/Diseases Health sciences/Health care Physical sciences/Mathematics and computing Health sciences/Medical research Machine Learning Patient Monitoring Predictive Analysis Time Series Clinical Support Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Medical complications in the general hospital ward, ranging from infection to heart failure occur in approximately 25% of all hospital admissions, thus affecting millions of patients annually [ 1 ], [ 2 ]. An unnoticed deterioration may lead to worsening of a medical complication, and a missed opportunity for intervention, which in turn might result in permanent injury, increased mortality, and raised healthcare expenses [ 3 ], [ 4 ]. This widespread problem highlights the importance of early deterioration detection to enable preventive measures that may reduce severity of an oncoming complication and ultimately improve outcomes for the patient [ 5 ], [ 6 ]. Currently, early prevention techniques at the general ward as a standard include intermittent manual measurements of vital signs including heart rate, blood oxygen saturation and respiration rate every 8–12 hours[ 7 ]. However, manual monitoring is labor-intensive and prone to inaccuracies due to human error, where complications might still arise in between measurements, potentially leaving the patient unobserved for 8–12 hours [ 8 ]. These factors explain why the most widely adopted intermittent manual monitoring system, the National Early Warning Score (NEWS and from here on referred to as the updated NEWS2) and related implementations, has failed to show impact on reducing morbidity and mortality remains [ 9 ]. To counter the inherent flaws of intermittent manual monitoring, continuous vital sign monitoring (CVSM) has gained international interest in general wards[ 10 ], [ 11 ], [ 12 ], [ 13 ]. CVSM’s main benefit is primarily greater sensitivity in detecting vital sign deviations compared to NEWS2, due to advancements in wearable technology and the continuous nature of the monitoring. If coupled to staff alerting CVSM allows for early interventions to reduce complication severity[ 14 ]. CVSM has recently been used to develop and implement alert systems based on the monitored vital signs, with significant success compared to NEWS2 [ 10 ]. Most CVSM-based alert systems are currently based on simple threshold values, and while this yields medical transparency and has improved patient outcomes, it does not fully leverage patient information or knowledge about past vital sign behavior [ 10 ], [ 15 ], [ 16 ]. Simple threshold based CVSM systems also pose clinical challenges due to a high rate of alerts, which can lead to alert fatigue among healthcare staff, leading them to discard the system[ 17 ], [ 18 ]. This is especially important in the context of the general hospital ward where one nurse may be responsible for 10 patients or more, rendering CVSM systems with hundreds of alerts per patients, clinically useless. Filtering alerts by requiring thresholds to be crossed for a given amount of time, reduces false alert counts[ 19 ]. Machine learning has the potential to improve monitoring of patients by extracting information from vital signs to identify early signs of deterioration related to complications [ 20 ], [ 21 ], [ 22 ], but is not being utilized in most patient monitoring systems for general wards. Missing values due to human error, patients removing equipment or technical errors further complicate the process and may lead to false alerts. Incorporating metadata into machine learning models may mitigate the lack of flexibility and adaptability in threshold-based systems and enable adjustment for individual patient characteristics. This paper aims to compare analyses of ML-based models versus state-of-the art threshold-and time-interval -based alert system to demonstrate the potential of ML in improving clinical monitoring and how different input data types and validation time frames impact detection performance. We hypothesize that an ML-driven system will significantly outperform a state-of-the-art threshold-based alert system by more accurately categorizing periods of vital signs as normal or problematic, while also overcoming the challenge posed by the frequent segments of missing data. Materials and Methods Data Description A dataset collected from 2423 hospitalized patients at Rigshospitalet and Bispebjerg Hospital, Copenhagen, Denmark, further described in the original publications [ 16 ], [ 23 ], [ 24 ], [ 25 ], was used for this study. The study was approved by the Danish Research Ethics Medical Committees (Videnskabsetiske Medicinske Komiteer). Each individual patient was asked to sign an informed consent statement separate from their treatment process. The patients were able to ask any follow up questions, consult relatives and staff, and finally stop participation in the data collection, at any given moment. All data collection studies adhered to the Helsinki declarations and had approval by local ethical committees. Each patient had the following data points: Continuous measurements: Heart rate, respiration rate, blood oxygen saturation, and pulse rate (recorded every minute). Semi-continuous measurements: Systolic blood pressure, diastolic blood pressure, and blood pressure-derived pulse rate. Recorded every 30 minutes in the daytime (0700–2200) and 60 minutes at night. 337 categories of potential medications received by the patients before hospitalization. 12 possible comorbidities. Each patient had an accompanying timestamp for all continuous and semi-continuous measurements. Distributions of vital sign values are shown in Fig. 1 .Vital sign data was recorded from three separate devices attached to patients as described in supplementary table 2. 1341 SAEs distributed across 865 patients were categorized as Cardiovascular (N = 402), Infectious (N = 188), Interventions (N = 236), Neurological (N = 80), Respiratory (N = 266), or Other (N = 169). These six categories were the result of grouping individual medical complication types, a process performed by a senior anaesthesiologist. These events served as the basis for evaluating alert relevance, with an ideal model triggering an alert for every complication and nowhere else. Data Preprocessing Vital sign data was matched with pre-existing patient comorbidities and medication histories from the Danish Electronic Healthcare System. A “Quality” measure was added based on the fraction of devices recording values at each time point and was later passed to the model. The dataset was split into a train (80%) and test (20%) set, ensuring that data from a single patient was never shared across the two datasets. Each SAE was used, even from patients with multiple SAEs. Missing data was handled by forward-filling to fill gaps of missing data and backfilling to close any gaps in the date at the beginning of the monitoring period. Figure 1 Distributions of Vital Sign Data Each patient’s data was divided into unique, non-overlapping 8 hour or 24-hour intervals, ensuring no two intervals shared data. All intervals were extended two hours beyond the final timestamp to account for clinical documentation delays and post-event vital sign fluctuations. For each 8-hour interval, gradients were calculated by linear regression every 60 minutes for each vital sign using only the vital signs in that specific interval. These gradient values along with the timestamp and intercept were then used to replace the vital sign data in that interval, effectively summarizing the greatest change found for a vital sign in that time interval. This process was repeated for the 24-hour intervals. Patient metadata (medication, comorbidities and demographic information) was encoded and included as input variables. Model Creation To predict a probability of a interval containing one or several SAE, a boosted decision tree-based[ 26 ] machine learning model BDT framework was made using XGBoost[ 27 ] 3.0.0. coded in Python 3.11. Data Augmentation To expand the dataset, artificial patient data was created by recombining measurements from different patients and devices. This process produced a total of 1500 additional intervals for both 8 hour and 24 hour durations for training. As illustrated in Fig. 2 , this process was divided into five for cross-validation intervals. The motivation for dividing the data into 8- and 24-hour intervals respectively stemmed from two factors: First, there was an unknown degree of uncertainty regarding the exact timestamp of the medical complications, due to their being recorded by healthcare staff after the complication itself occurred. Thus, it was unclear whether the vital signs recorded during those intervals captured all the relevant measurements needed to reflect the complication. Second, the duration length was chosen based on the regular length of a work shift, 8 hours, with 2 hours added for the uncertainty described in the methods section, or for the length of a full day, again with 2 hours added. This process was only done for training the model. Modelling For each split in the cross-validation 4 folds were used to train and test a separate BDT model, which was evaluated on the fifth and final fold. This method ensured fair performance comparison between models. By rotating which subset serves as the test set in each fold, cross-validation provided a more reliable estimate of model generalizability, while mitigating the risk of performance metrics being skewed by any single data split. Preventing patient overlap across folds eliminated information leakage, ensuring that each evaluation reflected the model’s true ability to generalize to unseen patients. Finally using cross-validation allowed us to measure the error of our model’s predictive power, with a high statistical certainty. True negative and false negative rates were also computed for model evaluation. Model performance on individual SAE types was evaluated by using the final models, which were trained with all SAE types, to only predict for each individual SAE type. Calling Alerts Threshold-based alerts were assigned based on predefined vital sign thresholds following those used in Grønbæk et al. 2023[ 28 ]. Each threshold alert was triggered by a sequence of either one or several values falling below a pre-defined threshold. Machine learning-based alerts were generated using the BDT, analyzing processed vital signs over 8-hour and 24-hour intervals. The ML model assigned a probability score to each interval, determining the likelihood of a medical complication. Validation of Alerts Alerts triggered within an interval containing a SAE were marked as True Positives, while those occurring outside of complication intervals were classified as False Positives. Alerts were validated by defining true positive intervals (intervals correctly predicting an event) and false positive intervals (incorrect alerts outside event windows). Comparing FPR and TPR To compare the FPR and TPR for the ML models to those from the threshold-based alerts, the FPR was fixed for the ML model based on the FPR from the corresponding threshold alerts (meaning same timeframe). From the fixed FPR it was possible to find the point on the models AUROC curve and thus a [FPR, TPR] coordinate set. The same method was used for TPR, fixing that instead. This allowed direct comparison of the difference in performance between the ML models and threshold-based alerts with either the same FPR or TPR. Results A total of 2325 patients were included in the analysis, after 98 patients were excluded due to having less than 8 hours of data or more than 14 days. Patient demographics are shown in table 1. Patients had a median age of 70 years, 40.4% were female. The main reason for medical admission was cancer surgery (20.3%) followed by abdominal surgery (17.5%). The 95% CI monitoring duration for the patients was 442–10173 minutes, with a mean monitoring duration 4862 minutes. As seen in Fig. 1 , the patient populations’ vital signs have wide distributions, except for oxygen saturation, which tended strongly toward the upper bound of 100%. The illustrated data was used to train our boosted decision tree-based machine learning model (BDT), shown in Fig. 2 . The ML model consistently outperformed the threshold-based alert system in identifying periods of medical deterioration based on vital sign data. Integration of patient metadata only marginally improved Area Under Receiver Operating Curve (AUROC), with 24 hour and 8 hour-based models showing an improvement in AUROC of + 0.01 and + 0.02 respectively. As illustrated in Fig. 3 , models trained on 24-hour windows demonstrated significantly greater predictive precision (AUROC = 0.73, 95% CI [0.72–0.74]) than those using 8-hour windows (AUROC = 0.67, 95% CI [0.64–0.70]), benefiting from a broader clinical context. Both approaches significantly outperformed the threshold-based system, with models incorporating patient-specific data further improving classification accuracy. The threshold-based system proved insufficient at predicting SAE in the 24-hour threshold and 8- hour threshold as shown in table 2 and Fig. 3 . As shown in Fig. 3 and in tables 2 and 3, the main findings indicated that a BDT consistently outperformed the threshold-based system for predicting medical complications, with AUROC scores for all models significantly outperforming thresholds-based systems as shown in table 3. When combining continuous vital sign data with patient-specific information such as comorbidities and medications, the BDT achieved better TPR and FPR, but only by 0.01–0.02. When comparing different time intervals, 24-hour windows were largely more effective than 8-hour windows, achieving higher TPR, with lower FPRs across models. The 24-hour interval provided more comprehensive information about the patient’s condition, improving the model’s ability to distinguish between complication and non-complication intervals. 24-hour models with metadata also proved significantly better (p < 10 − 100 ) from their threshold-based counterparts. Figure 4 Predictive performance on distinct event groups and variable contributions To further investigate on our results, we investigated model performance across six categories of medical deterioration, each of which exhibited distinct classification characteristics. Performance, as demonstrated in Fig. 4 , varied substantially when limiting the type of medical deterioration. As seen in Fig. 4 , the AUROC varied by up to ~ 0.2 across event types. The 24-hour ML model performed best in identifying most complications, except for cardiovascular SAEs where the 8-hour model proved slightly superior as shown in Fig. 4 a. Additionally, supplementary Fig. 1 highlighted the impact of removing specific monitoring devices (electrocardiogram, oximeter, and blood pressure monitor) from the model. Performance declined only by 0.01–0.02 AUROC when each monitoring device was omitted, suggesting that while all three contributed to prediction accuracy, no single device dominated overall model effectiveness. To statistically evaluate the ML model’s improvements over existing alerting systems, we conducted Fisher’s Exact Test, the results of which are summarized in table 2. The test compared the contingency tables, consisting of false positives, true positives, false negatives, and true negatives, between all variations of the ML model and both the threshold-based alert system and a random alert system. The p-values described the degree of significance with which the compared methods differed. As shown in table 2, the ML models demonstrated statistically significant differences from random alerting (p < 10 − 30 in all cases), confirming that predictions were very far from random. Discussion We hypothesized that machine learning could outperform threshold-based alerts. Our findings indicated that machine learning significantly outperformed a state-of-the-art threshold-severity based vital sign alert system [ 14 ] regarding forecasting oncoming SAEs, thus supporting our hypothesis. Furthermore, the improved CVSM alert precision was improved while potentially mitigating alert fatigue by reducing false alerts. Model improvements were especially improved by data augmentation and inclusion of real patient data. This study describes predictive performance of ML versus traditional threshold-based alert systems in predicting medical complications from CVSM data. As seen in table 3 the ML model vastly improved both True Positive Rate (TPR) and False Positive Rate (FPR), reinforcing its potential to enhance clinical monitoring while reducing alert fatigue. Prior work shown in Amet et. Al 2020 [ 29 ] and Logothetis et al. 2023 [ 21 ], suggests that ML algorithms are superior in predictive capability over NEWS-based monitor systems using CSVM. Recently continuous monitoring has proved effective in the ICU setting [ 30 ]. Our work shows that ML-based alerting algorithms have the capacity to improve threshold-based alert algorithms that are more advanced that NEWS-based monitoring. These findings are part of a growing shift toward data-driven approaches in hospital settings, where early detection of deterioration is critical for improving patient outcomes. A key result was the superior performance of 24-hour monitoring windows compared to 8-hour intervals. A longer observation period allowed the model to capture a broader physiological context, distinguishing meaningful clinical deteriorations from short-term variations. Therefore, the 24-hour window proved more effective in identifying critical changes, highlighting the importance of extended monitoring in predictive modelling. The lack of rapid-onset short-term SAEs likely contributed to the superior performance of 24-hour prediction windows. However, this manner of long-term predicting was not what the threshold-based system was designed for, which also explains the threshold systems limited performance in that temporal context as shown in Fig. 3 . Furthermore, integrating patient-specific metadata, including comorbidities and medication history, only slightly enhanced model performance for the 24-hour vital sign windows from AUROC = 0.72 to 0.73. This suggests that the model largely predicting based on vital sign behaviour and not only on patient metadata. Traditional threshold-based systems applied uniform criteria across all patients, failing to account for individual variability. In contrast, ML-based methods adapted dynamically, tailoring predictions to patient-specific risk factors, which might explain why including patient data slightly changed the predictive performance of the models. These findings highlighted the need for continuous monitoring strategies in clinical practice, which could significantly improve the specificity of early warning systems and reduce the false positive rate and resulting alert fatigue. An advantage of a ML model such as the BDT described here or more advanced models, is that they would allow for the calculation of a continuous risk score, which might be used in the threshold-based system to adjust vital sign thresholds, thus performing a running fine-tuning of alerts on an individual patient basis. Despite these promising results, several perspectives must be acknowledged. First, the dataset’s size and composition may limit generalizability. The study primarily analyses patients following major surgery or during acute medical treatment, which may not fully represent broader hospital populations. This suggests that individual models or fine-tuning of a larger model may be necessary to differentiate between patient types. Future research should incorporate larger, more diverse datasets across multiple institutions to validate the model’s robustness across populations. While BDT models are generally robust in their handling of missing data, exploring further the methods for handling missing data is also necessary, to fully capture potential contextual information in missing data, ie. is the data missing because the patient has taken off the devices while sleeping or because they are ambulatory. Second, while the ML model demonstrates superior predictive accuracy, real-world clinical integration remained an open challenge. Implementing ML-based alert systems requires seamless incorporation into existing hospital workflows, staff training, and willingness to implement new systems. Clinical trials are essential to assess not only predictive performance but also the practical impact on patient outcomes, resource allocation, and healthcare staff workload. From a broader perspective, these findings contributed to the ongoing conversation on ML-driven decision support in predictive healthcare. Prior studies demonstrated ML’s potential in outcome prediction based on categorical health record data. This study built on that foundation by demonstrating its effectiveness in general patient monitoring, suggesting a scalable approach for hospital-wide implementation. This study confirms that machine learning can significantly enhance CVSM-based early warning systems, outperforming traditional threshold-based alerts in identifying vital sign behaviour linked to medical deterioration. The use of longer monitoring windows and patient-specific metadata contributed to improved predictive accuracy, supporting the integration of personalized risk assessment into hospital monitoring. Cross-validation provided a robust method of confirming the enhanced predictive performance offered by ML models. Recombination of patient data, without SAE, for training purposes allowed an expansion the training dataset. Reliable ability to predict SAE in an interval was shown to be varied depending on the type of SAE. While clinical validation is still required, these findings indicate that ML-driven alerting systems have the potential to improve patient safety by enabling earlier, more accurate detection of clinical deterioration. Data Availability The datasets analyzed during the current study are not publicly available due to third party agreements (Capitol Region of Denmark), but are available from the senior author (EKA) upon reasonable request. Declarations Competing Interests CM and EA have created a spin-out company, WARD247 ApS, with the aim of pursuing the regulatory and commercial activities of the WARD-project (Wireless Assessment of Respiratory and circulatory Distress, a project developing a clinical support system for continuous wireless monitoring of vital signs). WARD247 ApS has obtained license agreement for any WARD-project software and patents. One patent has been filed: “Wireless Assessment of Respiratory and circulatory Distress (WARD), EP 21184712.4 and EP 21205557.8” EA reports institutional research funding from Norpharma A/S as well as lecture fees from Radiometer, and advisory roles for Concentric analgesics and GenEdit without relation to the present work. KG is currently employed as CMO at the spin-out company, WARD247 ApS. NP is currently employed as an industrial PhD at WARD247 ApS as a collaboration between the University of Copenhagen and the company. Competing Interest Statement: CSM and EKA have created a spin-out company, WARD247 ApS, with the aim of pursuing the regulatory and commercial activities of the WARD-project (Wireless Assessment of Respiratory and circulatory Distress, a project developing a clinical support system for continuous wireless monitoring of vital signs). WARD247 ApS has obtained license agreement for any WARD-project software and patents. One patent has been filed: “Wireless Assessment of Respiratory and circulatory Distress (WARD), EP 21184712.4 and EP 21205557.8” EKA reports institutional research funding from Norpharma A/S as well as lecture fees from Radiometer, and advisory roles for Concentric analgesics and GenEdit without relation to the present work. KG is currently employed as CMO at the spin-out company, WARD247 ApS. NKP is currently employed as an industrial PhD at WARD247 ApS as a collaboration between the University of Copenhagen and the company. Funding Declaration The authors would like to acknowledge the Danish Innovation Fund for providing substantial funding for both data collection and sponsoring Norman Pedersens PhD. Author Contribution Norman Pedersen: project administration, investigation, formal analysis, coding, writing. Andreas S. Hjortsø: coding and analysis. Carl I. Askehave: coding and analysis. Jesper Mølgaard: Data acquisition. Katja K. Grønbæk: Data acquisition and resources. Christian S. Meyhoff: Data acquisition, writing - reviewing and editing. Troels C. Petersen: Conceptualization, methodology, writing – reviewing and editing. Eske K. Aasvang: Funding conceptualization, methodology, writing – reviewing and editing. Data Availability The datasets analyzed during the current study are not publicly available due to third party agreements (Capitol Region of Denmark), but are available from the senior author (EKA) upon reasonable request. References Ghaferi, A. A., Birkmeyer, J. D. & Dimick, J. B. Variation in hospital mortality associated with inpatient surgery, N Engl J Med , vol. 361, no. 14, pp. 1368–1375, Oct. (2009). 10.1056/NEJMsa0903048 International Surgical Outcomes Study group. Global patient outcomes after elective surgery: prospective cohort study in 27 low-, middle- and high-income countries. Br. J. Anaesth. 117 (5), 601–609. 10.1093/bja/aew316 (Oct. 2016). Hoonhout, L. H. F. et al. Direct medical costs of adverse events in Dutch hospitals. BMC Health Serv. Res. 9 , 27. 10.1186/1472-6963-9-27 (Feb. 2009). Larsen, A. T. et al. Estimation of the maximum potential cost saving from reducing serious adverse events in hospitalized patients, Acta Anaesthesiol Scand , vol. 68, no. 10, pp. 1471–1480, Nov. (2024). 10.1111/aas.14525 Moonesinghe, S. R. et al. Survival after postoperative morbidity: a longitudinal observational cohort study. Br. J. Anaesth. 113 (6), 977–984. 10.1093/bja/aeu224 (Dec. 2014). Im, Y. et al. Time-to-antibiotics and clinical outcomes in patients with sepsis and septic shock: a prospective nationwide multicenter cohort study. Crit. Care . 26 (1), 19. 10.1186/s13054-021-03883-0 (Dec. 2022). National Early Warning Score (NEWS). 2. Accessed: Apr. 29, 2025. [Online]. Available: https://www.rcp.ac.uk/improving-care/resources/national-early-warning-score-news-2/ Briggs, J. et al. Mar., Safer and more efficient vital signs monitoring protocols to identify the deteriorating patients in the general hospital ward: an observational study, Health Soc Care Deliv Res , pp. 1–143, (2024). 10.3310/HYTR4612 Clifton, D. A. et al. Errors’ and omissions in paper-based early warning scores: the association with changes in vital signs–a database analysis. BMJ Open. 5 (7), e007376. 10.1136/bmjopen-2014-007376 (July 2015). Rowland, B. A., Motamedi, V., Michard, F., Saha, A. K. & Khanna, A. K. Impact of continuous and wireless monitoring of vital signs on clinical outcomes: a propensity-matched observational study of surgical ward patients, Br J Anaesth , vol. 132, no. 3, pp. 519–527, Mar. (2024). 10.1016/j.bja.2023.11.040 Dykes, P. C. et al. Reducing ICU Utilization, Length of Stay, and Cost by Optimizing the Clinical Use of Continuous Monitoring System Technology in the Hospital, Am J Med , vol. 135, no. 3, pp. 337–341.e1, Mar. (2022). 10.1016/j.amjmed.2021.09.024 Eddahchouri, Y. et al. Effect of continuous wireless vital sign monitoring on unplanned ICU admissions and rapid response team calls: a before-and-after study. Br. J. Anaesth. 128 (5), 857–863. 10.1016/j.bja.2022.01.036 (May 2022). Areia, C. et al. The impact of wearable continuous vital sign monitoring on deterioration detection and clinical outcomes in hospitalised patients: a systematic review and meta-analysis. Crit. Care . 25 (1), 351. 10.1186/s13054-021-03766-4 (Sept. 2021). Mølgaard, J. et al. Oct., Continuous Vital Sign Monitoring at the Surgical Ward for Improved Outcomes After Major Noncardiac Surgery: A Randomized Clinical Trial, Anesth Analg , vol. 141, no. 4, pp. 807–817, (2025). 10.1213/ANE.0000000000007606 Æ et al. Prediction of serious outcomes based on continuous vital sign monitoring of high-risk patients. Comput. Biol. Med. 147 , 105559. 10.1016/j.compbiomed.2022.105559 (Aug. 2022). Haahr-Raunkjaer, C. et al. Continuous monitoring of vital sign abnormalities; association to clinical complications in 500 postoperative patients. Acta Anaesthesiol. Scand. 66 (5), 552–562. 10.1111/aas.14048 (2022). Ruppel, H., Dougherty, M., Bonafide, C. P. & Lasater, K. B. Alarm burden and the nursing care environment: a 213-hospital cross-sectional study, BMJ Open Qual , vol. 12, no. 4, p. e002342, Oct. (2023). 10.1136/bmjoq-2023-002342 Sendelbach, S. & Funk, M. Alarm fatigue: a patient safety concern. AACN Adv. Crit. Care , 24 , 4, pp. 378–386 ; quiz 387–388, 2013, 10.1097/NCI.0b013e3182a903f9 Kjærgaard, K., Mølgaard, J., Rasmussen, S. M., Meyhoff, C. S. & Aasvang, E. K. The effect of technical filtering and clinical criteria on alert rates from continuous vital sign monitoring in the general ward, Hospital Practice , vol. 51, no. 5, pp. 295–302, Oct. (2023). 10.1080/21548331.2023.2298185 Kause, J. et al. A comparison of antecedents to cardiac arrests, deaths and emergency intensive care admissions in Australia and New Zealand, and the United Kingdom–the ACADEMIA study. Resuscitation 62 (3), 275–282. 10.1016/j.resuscitation.2004.05.016 (Sept. 2004). Boulitsakis Logothetis, S., Green, D., Holland, M. & Al Moubayed, N. Predicting acute clinical deterioration with interpretable machine learning to support emergency care decision making. Sci. Rep. 13 (1), 13563. 10.1038/s41598-023-40661-0 (Aug. 2023). Aldhoayan, M. D. & Aljubran, Y. Prediction of ICU Patients’ Deterioration Using Machine Learning Techniques. Cureus 15 (5), e38659. 10.7759/cureus.38659 (May 2023). Carlsson, C. J. et al. Apr., Continuous Glucose Monitoring Reveals Perioperative Hypoglycemia in Most Patients With Diabetes Undergoing Major Surgery: A Prospective Cohort Study, Ann Surg , vol. 277, no. 4, pp. 603–611, (2023). 10.1097/SLA.0000000000005246 Elvekjaer, M. et al. Sept., Clinical impact of vital sign abnormalities in patients admitted with acute exacerbation of chronic obstructive pulmonary disease: an observational study using continuous wireless monitoring, Intern Emerg Med , vol. 17, no. 6, pp. 1689–1698, (2022). 10.1007/s11739-022-02988-w Mølgaard, J. et al. Continuous wireless pre- and postoperative vital sign monitoring reveal new, severe desaturations after vascular surgery. Acta Anaesthesiol. Scand. 67 (1), 19–28. 10.1111/aas.14158 (2023). Roe, B. P. et al. Boosted decision trees, an alternative to artificial neural networks. Nucl. Instrum. Meth A . 543 , 2–3. 10.1016/j.nima.2004.12.018 (2005). Chen, T. & Guestrin, C. XGBoost: A Scalable Tree Boosting System, in Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , Aug. pp. 785–794. (2016). 10.1145/2939672.2939785 Grønbæk, K. K. et al. Continuous monitoring is superior to manual measurements in detecting vital sign deviations in patients with COVID-19. Acta Anaesthesiol. Scand. 67 (5), 640–648. 10.1111/aas.14221 (2023). Youssef Ali Amer, A. et al. Jan., Vital Signs Prediction and Early Warning Score Calculation Based on Continuous Monitoring of Hospitalised Patients Using Wearable Technology, Sensors , vol. 20, no. 22, Art. no. 22, (2020). 10.3390/s20226593 He, R. & Chiang, J. N. Simultaneous forecasting of vital sign trajectories in the ICU. Sci. Rep. 15 (1), 14996. 10.1038/s41598-025-99719-w (Apr. 2025). Tables Table 1: Patient demography. Average stay duration and demographic information for the patient dataset as well as threshold alert count and most common admission reason. Table 2: p-values for ML models tested against Random and Threshold-based alerts. An overview for the derived p-values from a Fishers Exact test for the BDT model. ML models are all significantly different from a random alert calling system and the threshold-based system. Table 3: Area Under Receiver Operating Curve with 95 % CI, False Positive Rate and True Positive Rates for all methods of alerting and models. Overview showing the calculated AUROC, FPR and TPR for 24-hour and 8-hour models with and without metadata. * Fixed FPR was calculated by using the FPR for the threshold alerts and finding the corresponding [FPR, TPR] coordinate on the AUROC curve for the appropriate model. **Finding the TPR used an identical method to * but using a fixed TPR value instead. Additional Declarations Competing interest reported. CM and EA have created a spin-out company, WARD247 ApS, with the aim of pursuing the regulatory and commercial activities of the WARD-project (Wireless Assessment of Respiratory and circulatory Distress, a project developing a clinical support system for continuous wireless monitoring of vital signs). WARD247 ApS has obtained license agreement for any WARD-project software and patents. One patent has been filed: “Wireless Assessment of Respiratory and circulatory Distress (WARD), EP 21184712.4 and EP 21205557.8” EA reports institutional research funding from Norpharma A/S as well as lecture fees from Radiometer, and advisory roles for Concentric analgesics and GenEdit without relation to the present work. KG is currently employed as CMO at the spin-out company, WARD247 ApS. NP is currently employed as an industrial PhD at WARD247 ApS as a collaboration between the University of Copenhagen and the company. Supplementary Files SupplementaryInformation.docx Cite Share Download PDF Status: Under Revision Version 1 posted Editorial decision: Revision requested 18 Feb, 2026 Reviews received at journal 18 Feb, 2026 Reviewers agreed at journal 09 Feb, 2026 Reviewers agreed at journal 06 Feb, 2026 Reviews received at journal 29 Jan, 2026 Reviewers agreed at journal 21 Jan, 2026 Reviewers agreed at journal 20 Jan, 2026 Reviewers agreed at journal 16 Jan, 2026 Reviewers agreed at journal 13 Jan, 2026 Reviewers invited by journal 13 Jan, 2026 Editor assigned by journal 12 Jan, 2026 Editor invited by journal 07 Jan, 2026 Submission checks completed at journal 05 Jan, 2026 First submitted to journal 05 Jan, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8436885","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":575414274,"identity":"442d13bc-2d61-4ce6-8872-a65a7295540a","order_by":0,"name":"Norman K. Pedersen","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABBElEQVRIiWNgGAWjYHCDBDApxyDB2MDYwMDAOINYLcaka0lskAAqx6fFnL336YafOxjy+NmTj338UXEnfbt0cwPjzDYG2ZkN2LVY9hw3u9l7hqFYsudZ8myeM89yd8452MC4sY3BeDYOWwxupLHd4G1jSNxwI8eYmbHtcO6GG4kNjA+BIvPwaLn5F6hg/438z4w//x1ONyBGy22wLRI5zAy8DYcTwFqADkvE6bAzx9huy7ZJFEuceWbMzHPssOGGOwcbDs44J2GMy/sGx9vYbr5ts8njb09+zPij5rC8we32hw97ymxkZxzAYQ0ESCSgcIGKJfCqB4EEgipGwSgYBaNg5AIAfaFlVRURql4AAAAASUVORK5CYII=","orcid":"","institution":"University of Copenhagen","correspondingAuthor":true,"prefix":"","firstName":"Norman","middleName":"K.","lastName":"Pedersen","suffix":""},{"id":575414275,"identity":"75677c31-110e-4472-8c18-8149fc0632a1","order_by":1,"name":"Andreas Hasselriis","email":"","orcid":"","institution":"University of Copenhagen","correspondingAuthor":false,"prefix":"","firstName":"Andreas","middleName":"","lastName":"Hasselriis","suffix":""},{"id":575414277,"identity":"d724a8db-594e-4899-8ad2-35df7c6b1400","order_by":2,"name":"Carl I. Askehave","email":"","orcid":"","institution":"University of Copenhagen","correspondingAuthor":false,"prefix":"","firstName":"Carl","middleName":"I.","lastName":"Askehave","suffix":""},{"id":575414278,"identity":"309edd09-1193-498b-82ca-77693f0f4f64","order_by":3,"name":"Jesper Mølgaard","email":"","orcid":"","institution":"Copenhagen University Hospital, Rigshospitalet","correspondingAuthor":false,"prefix":"","firstName":"Jesper","middleName":"","lastName":"Mølgaard","suffix":""},{"id":575414279,"identity":"e34a25a1-4537-444a-84b5-7926852f3a4a","order_by":4,"name":"Katja K. Grønbæk","email":"","orcid":"","institution":"Ward 24/7 Aps","correspondingAuthor":false,"prefix":"","firstName":"Katja","middleName":"K.","lastName":"Grønbæk","suffix":""},{"id":575414280,"identity":"c88e15ab-c559-4fcd-9ca3-c1fd609e2419","order_by":5,"name":"Christian S. Meyhoff","email":"","orcid":"","institution":"Copenhagen University Hospital - Bispebjerg and Frederiksberg","correspondingAuthor":false,"prefix":"","firstName":"Christian","middleName":"S.","lastName":"Meyhoff","suffix":""},{"id":575414281,"identity":"89a8ba28-258a-4a65-a788-44fd91d336c2","order_by":6,"name":"Eske K. Aasvang","email":"","orcid":"","institution":"Copenhagen University Hospital, Rigshospitalet","correspondingAuthor":false,"prefix":"","firstName":"Eske","middleName":"K.","lastName":"Aasvang","suffix":""},{"id":575414282,"identity":"17210412-f383-42d3-967b-032ec4e8fcbe","order_by":7,"name":"Troels C. Petersen","email":"","orcid":"","institution":"University of Copenhagen","correspondingAuthor":false,"prefix":"","firstName":"Troels","middleName":"C.","lastName":"Petersen","suffix":""}],"badges":[],"createdAt":"2025-12-23 20:08:17","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8436885/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8436885/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":100662272,"identity":"f71efdfc-67be-40e7-8a7d-97a85fe4b8c8","added_by":"auto","created_at":"2026-01-20 08:57:18","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1585615,"visible":true,"origin":"","legend":"","description":"","filename":"naturescientificreportsJAN0526.docx","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/9e7791e792bec36be63bd58b.docx"},{"id":100662364,"identity":"07f63864-f4f4-4fed-ad10-da371056c210","added_by":"auto","created_at":"2026-01-20 08:59:49","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10013,"visible":true,"origin":"","legend":"","description":"","filename":"20c6346d93bf483b838172ec3877b7e0.json","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/74642252280762fd8e079001.json"},{"id":100662253,"identity":"75df3ed0-7c61-492b-9a61-4c770a9f801b","added_by":"auto","created_at":"2026-01-20 08:56:46","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3571682,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryInformation.docx","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/1036cc5032bf55ff05bac353.docx"},{"id":100662304,"identity":"61db3520-c2af-4aee-8f9a-30d442963c84","added_by":"auto","created_at":"2026-01-20 08:58:39","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":88344,"visible":true,"origin":"","legend":"","description":"","filename":"20c6346d93bf483b838172ec3877b7e01enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/2bf5409395e00cb504bd4de7.xml"},{"id":100662271,"identity":"62912bb7-7340-4c8c-8c54-d7361de148e1","added_by":"auto","created_at":"2026-01-20 08:57:17","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":141509,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/3286080d4d03d313640fad9d.png"},{"id":100662293,"identity":"1cf0b7e9-ba5f-4c07-95c5-c54ffddc49fd","added_by":"auto","created_at":"2026-01-20 08:58:12","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":86801,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/c8f8a5d1d54d0649e5c56dfb.png"},{"id":100662344,"identity":"fc2d4147-dc26-4f54-9aaf-e5ddea35bc1a","added_by":"auto","created_at":"2026-01-20 08:59:20","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":153513,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/aec33bf7a93bfceabfb77223.png"},{"id":100662275,"identity":"ebf2984e-a12e-4d4e-9b2c-1dcf7a8cfbd2","added_by":"auto","created_at":"2026-01-20 08:57:25","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":56275,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/39b788095c3b19f52db6aeae.png"},{"id":100662287,"identity":"c6ab1e50-8655-44c6-b822-08d296478beb","added_by":"auto","created_at":"2026-01-20 08:57:54","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":54282,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/474849275d0a1af419e47ffd.png"},{"id":100662312,"identity":"77f40f1d-3ed1-4378-9f05-1b4f313c8e21","added_by":"auto","created_at":"2026-01-20 08:58:44","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":61364,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/cca8a6e52e5798582d40cdd8.png"},{"id":100662255,"identity":"b871edf1-4646-4b33-a9da-350d27a33999","added_by":"auto","created_at":"2026-01-20 08:56:52","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":97365,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/7cd8eff9d9614b68f5b3aee2.png"},{"id":100662283,"identity":"5ee88953-5d91-4d2e-b84d-2ffe9d42eb3f","added_by":"auto","created_at":"2026-01-20 08:57:39","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":67008,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/09f35911f3bee42ddcbeb869.png"},{"id":100662289,"identity":"21804efe-0040-4b9b-a9c9-4451df8c77cd","added_by":"auto","created_at":"2026-01-20 08:58:05","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":50146,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/6542f525ca26eeca1399488c.png"},{"id":100662310,"identity":"2cae27d0-6463-470c-9676-c5e02ca59745","added_by":"auto","created_at":"2026-01-20 08:58:43","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":84201,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/9640fc53b821486a254be5b1.png"},{"id":100662292,"identity":"d30cac14-8cc8-44a4-b3eb-72e8cadc99b1","added_by":"auto","created_at":"2026-01-20 08:58:11","extension":"xml","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":83915,"visible":true,"origin":"","legend":"","description":"","filename":"20c6346d93bf483b838172ec3877b7e01structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/fd14c667d7c3b5288043528c.xml"},{"id":100662328,"identity":"ccab453e-484b-490d-8cb1-20842aab1e69","added_by":"auto","created_at":"2026-01-20 08:59:02","extension":"html","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":99166,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/514530bfd62029627309fc4c.html"},{"id":100662282,"identity":"87c06344-fe07-4ac4-969b-1cb9077c1e53","added_by":"auto","created_at":"2026-01-20 08:57:34","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":217586,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDistributions of Vital Sign Data\u003c/strong\u003e (\u003cem\u003ea) and (b), marked in blue, depict heart and respiration rate pr minute measured by a single lead ECG-patch attached on the chest of the patient. (c) and (d), marked in red, show pulse rate pr. min. and % blood oxygen saturation pr. min. measured by an Oximeter device attached to the patient’s finger. (e), (d) and (g), marked in green, represent the distributions of systolic and diastolic blood pressure as well as an alternative pulse rate all measured pr. ≈ 30 min. by a blood pressure cuff attached to the patient's arm.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/c542c595d59b5689a8f55d0c.png"},{"id":100662297,"identity":"1c29cb66-c18f-494d-a55e-0c96bbb9f0a7","added_by":"auto","created_at":"2026-01-20 08:58:22","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":270416,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eData Pipeline \u003c/strong\u003e\u003cem\u003eRaw data is first conjoined with complication data and patient metadata. Each patient’s data is then divided into intervals of a fixed length. For each interval, the maximum slope, intercept, and time for each vital sign are calculated. These values are added to the dataset, and the raw vital signs are removed. The data is then split 80% / 20% into training and test sets, with the test set siloed off. The training set is further augmented by creating synthetic patients through replacing one of the three devices’ measurements with data from another patient, thereby expanding the training dataset. The augmented training dataset is used to train a boosted decision tree model, which is ultimately used to produce a probability for each interval in the test set, indicating the likelihood that the interval contains at least one SAE.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/5ab52739f7869f7fa6f9759a.png"},{"id":100662294,"identity":"b37f1afc-5fdf-4ff8-8dee-506ae1bcb91d","added_by":"auto","created_at":"2026-01-20 08:58:13","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":197110,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePerformance of Machine Learning Models \u003c/strong\u003e\u003cem\u003eThis figure shows mean AUROC curves with 1σ error bands for four models, comparing True Positive Rate (TPR) versus False Positive Rate (FPR). The four configurations, all with slopes calculated over 60 minutes, include: A 24-hour window using One-Hot-Encoding (OHE) (blue), A 24-hour window without OHE (blue dotted), A 8-hour window with OHE (orange), A 8-hour window without OHE (orange dotted). Shaded areas represent the 1σ uncertainty around each curve. Purple and brown dots indicate the performance of current threshold-based alerting systems for the 24-hour and 8-hour models, respectively. Models using encoded metadata generally show slightly improved performance, with higher TPRs for a given FPR compared to models without encoding.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/6df58e58e6a74e187a9c5c2d.png"},{"id":100662290,"identity":"7f6bab12-f2b5-438d-a96f-bfd23ae9abfa","added_by":"auto","created_at":"2026-01-20 08:58:08","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":475436,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePredictive performance on distinct event groups and variable contributions \u003c/strong\u003e\u003cem\u003eThis figure shows the predictive performance of the machine learning model in sub-figures A-E on five distinct groups of event types: Cardiovascular (a), Infectious (b), Interventions (c), Neurological (d), Other(e) and Respiratory(f). AUROC varies by up to 0.1, indicating significant differences in the model’s ability to correctly classify an interval depending on the type of event present in that period.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/c62629320526c18277279580.png"},{"id":100665788,"identity":"fa8fa1c5-09f4-47b9-abb1-9b7ddae39e8d","added_by":"auto","created_at":"2026-01-20 09:30:26","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2075577,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/d248fefb-7a8e-4d24-b901-c33c66f1e718.pdf"},{"id":100662295,"identity":"6b4a3414-eaeb-4385-9b31-527b022cfe15","added_by":"auto","created_at":"2026-01-20 08:58:14","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":3571682,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryInformation.docx","url":"https://assets-eu.researchsquare.com/files/rs-8436885/v1/55c51d17028b5b78c78bc192.docx"}],"financialInterests":"Competing interest reported. CM and EA have created a spin-out company, WARD247 ApS, with the aim of pursuing the regulatory and commercial activities of the WARD-project (Wireless Assessment of Respiratory and circulatory Distress, a project developing a clinical support system for continuous wireless monitoring of vital signs). WARD247 ApS has obtained license agreement for any WARD-project software and patents. One patent has been filed: “Wireless Assessment of Respiratory and circulatory Distress (WARD), EP 21184712.4 and EP 21205557.8” EA reports institutional research funding from Norpharma A/S as well as lecture fees from Radiometer, and advisory roles for Concentric analgesics and GenEdit without relation to the present work. KG is currently employed as CMO at the spin-out company, WARD247 ApS. NP is currently employed as an industrial PhD at WARD247 ApS as a collaboration between the University of Copenhagen and the company.","formattedTitle":"Machine learning outperforms state-of-the-art continuous vital sign monitoring","fulltext":[{"header":"Introduction","content":"\u003cp\u003eMedical complications in the general hospital ward, ranging from infection to heart failure occur in approximately 25% of all hospital admissions, thus affecting millions of patients annually [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e], [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. An unnoticed deterioration may lead to worsening of a medical complication, and a missed opportunity for intervention, which in turn might result in permanent injury, increased mortality, and raised healthcare expenses [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e], [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. This widespread problem highlights the importance of early deterioration detection to enable preventive measures that may reduce severity of an oncoming complication and ultimately improve outcomes for the patient [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e], [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eCurrently, early prevention techniques at the general ward as a standard include intermittent manual measurements of vital signs including heart rate, blood oxygen saturation and respiration rate every 8\u0026ndash;12 hours[\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. However, manual monitoring is labor-intensive and prone to inaccuracies due to human error, where complications might still arise in between measurements, potentially leaving the patient unobserved for 8\u0026ndash;12 hours [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. These factors explain why the most widely adopted intermittent manual monitoring system, the National Early Warning Score (NEWS and from here on referred to as the updated NEWS2) and related implementations, has failed to show impact on reducing morbidity and mortality remains [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTo counter the inherent flaws of intermittent manual monitoring, continuous vital sign monitoring (CVSM) has gained international interest in general wards[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e], [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e], [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. CVSM\u0026rsquo;s main benefit is primarily greater sensitivity in detecting vital sign deviations compared to NEWS2, due to advancements in wearable technology and the continuous nature of the monitoring. If coupled to staff alerting CVSM allows for early interventions to reduce complication severity[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. CVSM has recently been used to develop and implement alert systems based on the monitored vital signs, with significant success compared to NEWS2 [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Most CVSM-based alert systems are currently based on simple threshold values, and while this yields medical transparency and has improved patient outcomes, it does not fully leverage patient information or knowledge about past vital sign behavior [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e], [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e], [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e]. Simple threshold based CVSM systems also pose clinical challenges due to a high rate of alerts, which can lead to alert fatigue among healthcare staff, leading them to discard the system[\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e], [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. This is especially important in the context of the general hospital ward where one nurse may be responsible for 10 patients or more, rendering CVSM systems with hundreds of alerts per patients, clinically useless. Filtering alerts by requiring thresholds to be crossed for a given amount of time, reduces false alert counts[\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Machine learning has the potential to improve monitoring of patients by extracting information from vital signs to identify early signs of deterioration related to complications [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e], [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e], but is not being utilized in most patient monitoring systems for general wards. Missing values due to human error, patients removing equipment or technical errors further complicate the process and may lead to false alerts. Incorporating metadata into machine learning models may mitigate the lack of flexibility and adaptability in threshold-based systems and enable adjustment for individual patient characteristics.\u003c/p\u003e \u003cp\u003eThis paper aims to compare analyses of ML-based models versus state-of-the art threshold-and time-interval -based alert system to demonstrate the potential of ML in improving clinical monitoring and how different input data types and validation time frames impact detection performance. We hypothesize that an ML-driven system will significantly outperform a state-of-the-art threshold-based alert system by more accurately categorizing periods of vital signs as normal or problematic, while also overcoming the challenge posed by the frequent segments of missing data.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eData Description\u003c/h2\u003e \u003cp\u003eA dataset collected from 2423 hospitalized patients at Rigshospitalet and Bispebjerg Hospital, Copenhagen, Denmark, further described in the original publications [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e], [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e], [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], was used for this study. The study was approved by the Danish Research Ethics Medical Committees (Videnskabsetiske Medicinske Komiteer).\u003c/p\u003e \u003cp\u003e Each individual patient was asked to sign an informed consent statement separate from their treatment process. The patients were able to ask any follow up questions, consult relatives and staff, and finally stop participation in the data collection, at any given moment. All data collection studies adhered to the Helsinki declarations and had approval by local ethical committees.\u003c/p\u003e \u003cp\u003eEach patient had the following data points:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eContinuous measurements: Heart rate, respiration rate, blood oxygen saturation, and pulse rate (recorded every minute).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eSemi-continuous measurements: Systolic blood pressure, diastolic blood pressure, and blood pressure-derived pulse rate. Recorded every 30 minutes in the daytime (0700\u0026ndash;2200) and 60 minutes at night.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e337 categories of potential medications received by the patients before hospitalization.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e12 possible comorbidities.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eEach patient had an accompanying timestamp for all continuous and semi-continuous measurements. Distributions of vital sign values are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.Vital sign data was recorded from three separate devices attached to patients as described in supplementary table 2.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e1341 SAEs distributed across 865 patients were categorized as Cardiovascular (N\u0026thinsp;=\u0026thinsp;402), Infectious (N\u0026thinsp;=\u0026thinsp;188), Interventions (N\u0026thinsp;=\u0026thinsp;236), Neurological (N\u0026thinsp;=\u0026thinsp;80), Respiratory (N\u0026thinsp;=\u0026thinsp;266), or Other (N\u0026thinsp;=\u0026thinsp;169). These six categories were the result of grouping individual medical complication types, a process performed by a senior anaesthesiologist. These events served as the basis for evaluating alert relevance, with an ideal model triggering an alert for every complication and nowhere else.\u003c/p\u003e \u003cp\u003e \u003cb\u003eData Preprocessing\u003c/b\u003e \u003c/p\u003e \u003cp\u003e Vital sign data was matched with pre-existing patient comorbidities and medication histories from the Danish Electronic Healthcare System. A \u0026ldquo;Quality\u0026rdquo; measure was added based on the fraction of devices recording values at each time point and was later passed to the model. The dataset was split into a train (80%) and test (20%) set, ensuring that data from a single patient was never shared across the two datasets. Each SAE was used, even from patients with multiple SAEs. Missing data was handled by forward-filling to fill gaps of missing data and backfilling to close any gaps in the date at the beginning of the monitoring period.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e \u003cb\u003eDistributions of Vital Sign Data\u003c/b\u003e\u003c/p\u003e \u003cp\u003eEach patient\u0026rsquo;s data was divided into unique, non-overlapping 8 hour or 24-hour intervals, ensuring no two intervals shared data. All intervals were extended two hours beyond the final timestamp to account for clinical documentation delays and post-event vital sign fluctuations. For each 8-hour interval, gradients were calculated by linear regression every 60 minutes for each vital sign using only the vital signs in that specific interval. These gradient values along with the timestamp and intercept were then used to replace the vital sign data in that interval, effectively summarizing the greatest change found for a vital sign in that time interval. This process was repeated for the 24-hour intervals. Patient metadata (medication, comorbidities and demographic information) was encoded and included as input variables.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eModel Creation\u003c/h3\u003e\n\u003cp\u003eTo predict a probability of a interval containing one or several SAE, a boosted decision tree-based[\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e] machine learning model BDT framework was made using XGBoost[\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e] 3.0.0. coded in Python 3.11.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eData Augmentation\u003c/h3\u003e\n\u003cp\u003eTo expand the dataset, artificial patient data was created by recombining measurements from different patients and devices. This process produced a total of 1500 additional intervals for both 8 hour and 24 hour durations for training. As illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e2\u003c/span\u003e, this process was divided into five for cross-validation intervals. The motivation for dividing the data into 8- and 24-hour intervals respectively stemmed from two factors: First, there was an unknown degree of uncertainty regarding the exact timestamp of the medical complications, due to their being recorded by healthcare staff after the complication itself occurred. Thus, it was unclear whether the vital signs recorded during those intervals captured all the relevant measurements needed to reflect the complication. Second, the duration length was chosen based on the regular length of a work shift, 8 hours, with 2 hours added for the uncertainty described in the methods section, or for the length of a full day, again with 2 hours added. This process was only done for training the model.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e\n\u003ch3\u003eModelling\u003c/h3\u003e\n\u003cp\u003eFor each split in the cross-validation 4 folds were used to train and test a separate BDT model, which was evaluated on the fifth and final fold. This method ensured fair performance comparison between models. By rotating which subset serves as the test set in each fold, cross-validation provided a more reliable estimate of model generalizability, while mitigating the risk of performance metrics being skewed by any single data split. Preventing patient overlap across folds eliminated information leakage, ensuring that each evaluation reflected the model\u0026rsquo;s true ability to generalize to unseen patients. Finally using cross-validation allowed us to measure the error of our model\u0026rsquo;s predictive power, with a high statistical certainty. True negative and false negative rates were also computed for model evaluation. Model performance on individual SAE types was evaluated by using the final models, which were trained with all SAE types, to only predict for each individual SAE type.\u003c/p\u003e\n\u003ch3\u003eCalling Alerts\u003c/h3\u003e\n\u003cp\u003eThreshold-based alerts were assigned based on predefined vital sign thresholds following those used in Gr\u0026oslash;nb\u0026aelig;k et al. 2023[\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e]. Each threshold alert was triggered by a sequence of either one or several values falling below a pre-defined threshold. Machine learning-based alerts were generated using the BDT, analyzing processed vital signs over 8-hour and 24-hour intervals. The ML model assigned a probability score to each interval, determining the likelihood of a medical complication.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eValidation of Alerts\u003c/h2\u003e \u003cp\u003eAlerts triggered within an interval containing a SAE were marked as True Positives, while those occurring outside of complication intervals were classified as False Positives. Alerts were validated by defining true positive intervals (intervals correctly predicting an event) and false positive intervals (incorrect alerts outside event windows).\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eComparing FPR and TPR\u003c/h3\u003e\n\u003cp\u003eTo compare the FPR and TPR for the ML models to those from the threshold-based alerts, the FPR was fixed for the ML model based on the FPR from the corresponding threshold alerts (meaning same timeframe). From the fixed FPR it was possible to find the point on the models AUROC curve and thus a [FPR, TPR] coordinate set. The same method was used for TPR, fixing that instead. This allowed direct comparison of the difference in performance between the ML models and threshold-based alerts with either the same FPR or TPR.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eA total of 2325 patients were included in the analysis, after 98 patients were excluded due to having less than 8 hours of data or more than 14 days. Patient demographics are shown in table 1. Patients had a median age of 70 years, 40.4% were female. The main reason for medical admission was cancer surgery (20.3%) followed by abdominal surgery (17.5%). The 95% CI monitoring duration for the patients was 442\u0026ndash;10173 minutes, with a mean monitoring duration 4862 minutes.\u003c/p\u003e \u003cp\u003eAs seen in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, the patient populations\u0026rsquo; vital signs have wide distributions, except for oxygen saturation, which tended strongly toward the upper bound of 100%. The illustrated data was used to train our boosted decision tree-based machine learning model (BDT), shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e2\u003c/span\u003e. The ML model consistently outperformed the threshold-based alert system in identifying periods of medical deterioration based on vital sign data. Integration of patient metadata only marginally improved Area Under Receiver Operating Curve (AUROC), with 24 hour and 8 hour-based models showing an improvement in AUROC of +\u0026thinsp;0.01 and +\u0026thinsp;0.02 respectively. As illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003e, models trained on 24-hour windows demonstrated significantly greater predictive precision (AUROC\u0026thinsp;=\u0026thinsp;0.73, 95% CI [0.72\u0026ndash;0.74]) than those using 8-hour windows (AUROC\u0026thinsp;=\u0026thinsp;0.67, 95% CI [0.64\u0026ndash;0.70]), benefiting from a broader clinical context. Both approaches significantly outperformed the threshold-based system, with models incorporating patient-specific data further improving classification accuracy. The threshold-based system proved insufficient at predicting SAE in the 24-hour threshold and 8- hour threshold as shown in table 2 and Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003e and in tables 2 and 3, the main findings indicated that a BDT consistently outperformed the threshold-based system for predicting medical complications, with AUROC scores for all models significantly outperforming thresholds-based systems as shown in table 3. When combining continuous vital sign data with patient-specific information such as comorbidities and medications, the BDT achieved better TPR and FPR, but only by 0.01\u0026ndash;0.02.\u003c/p\u003e \u003cp\u003eWhen comparing different time intervals, 24-hour windows were largely more effective than 8-hour windows, achieving higher TPR, with lower FPRs across models. The 24-hour interval provided more comprehensive information about the patient\u0026rsquo;s condition, improving the model\u0026rsquo;s ability to distinguish between complication and non-complication intervals. 24-hour models with metadata also proved significantly better (p\u0026thinsp;\u0026lt;\u0026thinsp;10\u003csup\u003e\u0026minus;\u0026thinsp;100\u003c/sup\u003e) from their threshold-based counterparts.\u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e4\u003c/span\u003e \u003cb\u003ePredictive performance on distinct event groups and variable contributions\u003c/b\u003e\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo further investigate on our results, we investigated model performance across six categories of medical deterioration, each of which exhibited distinct classification characteristics. Performance, as demonstrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e4\u003c/span\u003e, varied substantially when limiting the type of medical deterioration. As seen in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e4\u003c/span\u003e, the AUROC varied by up to ~\u0026thinsp;0.2 across event types. The 24-hour ML model performed best in identifying most complications, except for cardiovascular SAEs where the 8-hour model proved slightly superior as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e4\u003c/span\u003ea. Additionally, supplementary Fig.\u0026nbsp;1 highlighted the impact of removing specific monitoring devices (electrocardiogram, oximeter, and blood pressure monitor) from the model. Performance declined only by 0.01\u0026ndash;0.02 AUROC when each monitoring device was omitted, suggesting that while all three contributed to prediction accuracy, no single device dominated overall model effectiveness.\u003c/p\u003e \u003cp\u003eTo statistically evaluate the ML model\u0026rsquo;s improvements over existing alerting systems, we conducted Fisher\u0026rsquo;s Exact Test, the results of which are summarized in table 2. The test compared the contingency tables, consisting of false positives, true positives, false negatives, and true negatives, between all variations of the ML model and both the threshold-based alert system and a random alert system. The p-values described the degree of significance with which the compared methods differed. As shown in table 2, the ML models demonstrated statistically significant differences from random alerting (p\u0026thinsp;\u0026lt;\u0026thinsp;10\u003csup\u003e\u0026minus;\u0026thinsp;30\u003c/sup\u003e in all cases), confirming that predictions were very far from random.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eWe hypothesized that machine learning could outperform threshold-based alerts. Our findings indicated that machine learning significantly outperformed a state-of-the-art threshold-severity based vital sign alert system [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e] regarding forecasting oncoming SAEs, thus supporting our hypothesis. Furthermore, the improved CVSM alert precision was improved while potentially mitigating alert fatigue by reducing false alerts. Model improvements were especially improved by data augmentation and inclusion of real patient data.\u003c/p\u003e \u003cp\u003eThis study describes predictive performance of ML versus traditional threshold-based alert systems in predicting medical complications from CVSM data. As seen in table 3 the ML model vastly improved both True Positive Rate (TPR) and False Positive Rate (FPR), reinforcing its potential to enhance clinical monitoring while reducing alert fatigue. Prior work shown in Amet et. Al 2020 [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e] and Logothetis et al. 2023 [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], suggests that ML algorithms are superior in predictive capability over NEWS-based monitor systems using CSVM. Recently continuous monitoring has proved effective in the ICU setting [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. Our work shows that ML-based alerting algorithms have the capacity to improve threshold-based alert algorithms that are more advanced that NEWS-based monitoring. These findings are part of a growing shift toward data-driven approaches in hospital settings, where early detection of deterioration is critical for improving patient outcomes.\u003c/p\u003e \u003cp\u003eA key result was the superior performance of 24-hour monitoring windows compared to 8-hour intervals. A longer observation period allowed the model to capture a broader physiological context, distinguishing meaningful clinical deteriorations from short-term variations. Therefore, the 24-hour window proved more effective in identifying critical changes, highlighting the importance of extended monitoring in predictive modelling. The lack of rapid-onset short-term SAEs likely contributed to the superior performance of 24-hour prediction windows. However, this manner of long-term predicting was not what the threshold-based system was designed for, which also explains the threshold systems limited performance in that temporal context as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eFurthermore, integrating patient-specific metadata, including comorbidities and medication history, only slightly enhanced model performance for the 24-hour vital sign windows from AUROC\u0026thinsp;=\u0026thinsp;0.72 to 0.73. This suggests that the model largely predicting based on vital sign behaviour and not only on patient metadata. Traditional threshold-based systems applied uniform criteria across all patients, failing to account for individual variability. In contrast, ML-based methods adapted dynamically, tailoring predictions to patient-specific risk factors, which might explain why including patient data slightly changed the predictive performance of the models. These findings highlighted the need for continuous monitoring strategies in clinical practice, which could significantly improve the specificity of early warning systems and reduce the false positive rate and resulting alert fatigue.\u003c/p\u003e \u003cp\u003eAn advantage of a ML model such as the BDT described here or more advanced models, is that they would allow for the calculation of a continuous risk score, which might be used in the threshold-based system to adjust vital sign thresholds, thus performing a running fine-tuning of alerts on an individual patient basis.\u003c/p\u003e \u003cp\u003eDespite these promising results, several perspectives must be acknowledged. First, the dataset\u0026rsquo;s size and composition may limit generalizability. The study primarily analyses patients following major surgery or during acute medical treatment, which may not fully represent broader hospital populations. This suggests that individual models or fine-tuning of a larger model may be necessary to differentiate between patient types. Future research should incorporate larger, more diverse datasets across multiple institutions to validate the model\u0026rsquo;s robustness across populations. While BDT models are generally robust in their handling of missing data, exploring further the methods for handling missing data is also necessary, to fully capture potential contextual information in missing data, ie. is the data missing because the patient has taken off the devices while sleeping or because they are ambulatory.\u003c/p\u003e \u003cp\u003eSecond, while the ML model demonstrates superior predictive accuracy, real-world clinical integration remained an open challenge. Implementing ML-based alert systems requires seamless incorporation into existing hospital workflows, staff training, and willingness to implement new systems. Clinical trials are essential to assess not only predictive performance but also the practical impact on patient outcomes, resource allocation, and healthcare staff workload.\u003c/p\u003e \u003cp\u003eFrom a broader perspective, these findings contributed to the ongoing conversation on ML-driven decision support in predictive healthcare. Prior studies demonstrated ML\u0026rsquo;s potential in outcome prediction based on categorical health record data. This study built on that foundation by demonstrating its effectiveness in general patient monitoring, suggesting a scalable approach for hospital-wide implementation.\u003c/p\u003e \u003cp\u003eThis study confirms that machine learning can significantly enhance CVSM-based early warning systems, outperforming traditional threshold-based alerts in identifying vital sign behaviour linked to medical deterioration. The use of longer monitoring windows and patient-specific metadata contributed to improved predictive accuracy, supporting the integration of personalized risk assessment into hospital monitoring. Cross-validation provided a robust method of confirming the enhanced predictive performance offered by ML models. Recombination of patient data, without SAE, for training purposes allowed an expansion the training dataset. Reliable ability to predict SAE in an interval was shown to be varied depending on the type of SAE.\u003c/p\u003e \u003cp\u003eWhile clinical validation is still required, these findings indicate that ML-driven alerting systems have the potential to improve patient safety by enabling earlier, more accurate detection of clinical deterioration.\u003c/p\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eData Availability\u003c/h2\u003e \u003cp\u003eThe datasets analyzed during the current study are not publicly available due to third party agreements (Capitol Region of Denmark), but are available from the senior author (EKA) upon reasonable request.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003ch2\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/h2\u003e\n\u003cp\u003eCM and EA have created a spin-out company, WARD247 ApS, with the aim of pursuing the regulatory and commercial activities of the WARD-project (Wireless Assessment of Respiratory and circulatory Distress, a project developing a clinical support system for continuous wireless monitoring of vital signs). WARD247 ApS has obtained license agreement for any WARD-project software and patents. One patent has been filed: \u0026ldquo;Wireless Assessment of Respiratory and circulatory Distress (WARD), EP 21184712.4 and EP 21205557.8\u0026rdquo; EA reports institutional research funding from Norpharma A/S as well as lecture fees from Radiometer, and advisory roles for Concentric analgesics and GenEdit without relation to the present work. KG is currently employed as CMO at the spin-out company, WARD247 ApS. NP is currently employed as an industrial PhD at WARD247 ApS as a collaboration between the University of Copenhagen and the company.\u003c/p\u003e\n\u003ch2\u003eCompeting Interest Statement:\u003c/h2\u003e\n\u003cp\u003eCSM and EKA have created a spin-out company, WARD247 ApS, with the aim of pursuing the regulatory and commercial activities of the WARD-project (Wireless Assessment of Respiratory and circulatory Distress, a project developing a clinical support system for continuous wireless monitoring of vital signs). WARD247 ApS has obtained license agreement for any WARD-project software and patents. One patent has been filed: \u0026ldquo;Wireless Assessment of Respiratory and circulatory Distress (WARD), EP 21184712.4 and EP 21205557.8\u0026rdquo; EKA reports institutional research funding from Norpharma A/S as well as lecture fees from Radiometer, and advisory roles for Concentric analgesics and GenEdit without relation to the present work. KG is currently employed as CMO at the spin-out company, WARD247 ApS. NKP is currently employed as an industrial PhD at WARD247 ApS as a collaboration between the University of Copenhagen and the company.\u003c/p\u003e\n\u003ch2\u003eFunding\u003c/h2\u003e\n\u003cp\u003eDeclaration\u003c/p\u003e\n\u003cp\u003eThe authors would like to acknowledge the Danish Innovation Fund for providing substantial funding for both data collection and sponsoring Norman Pedersens PhD.\u003c/p\u003e\n\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\n\u003cp\u003eNorman Pedersen: project administration, investigation, formal analysis, coding, writing. Andreas S. Hjorts\u0026oslash;: coding and analysis. Carl I. Askehave: coding and analysis. Jesper M\u0026oslash;lgaard: Data acquisition. Katja K. Gr\u0026oslash;nb\u0026aelig;k: Data acquisition and resources. Christian S. Meyhoff: Data acquisition, writing - reviewing and editing. Troels C. Petersen: Conceptualization, methodology, writing \u0026ndash; reviewing and editing. Eske K. Aasvang: Funding conceptualization, methodology, writing \u0026ndash; reviewing and editing.\u003c/p\u003e\n\u003ch2\u003eData Availability\u003c/h2\u003e\n\u003cp\u003eThe datasets analyzed during the current study are not publicly available due to third party agreements (Capitol Region of Denmark), but are available from the\u0026nbsp;senior\u0026nbsp;author (EKA) upon reasonable request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eGhaferi, A. A., Birkmeyer, J. D. \u0026amp; Dimick, J. B. Variation in hospital mortality associated with inpatient surgery, \u003cem\u003eN Engl J Med\u003c/em\u003e, vol. 361, no. 14, pp. 1368\u0026ndash;1375, Oct. (2009). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1056/NEJMsa0903048\u003c/span\u003e\u003cspan address=\"10.1056/NEJMsa0903048\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eInternational Surgical Outcomes Study group. Global patient outcomes after elective surgery: prospective cohort study in 27 low-, middle- and high-income countries. \u003cem\u003eBr. J. Anaesth.\u003c/em\u003e \u003cb\u003e117\u003c/b\u003e (5), 601\u0026ndash;609. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/bja/aew316\u003c/span\u003e\u003cspan address=\"10.1093/bja/aew316\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (Oct. 2016).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHoonhout, L. H. F. et al. Direct medical costs of adverse events in Dutch hospitals. \u003cem\u003eBMC Health Serv. Res.\u003c/em\u003e \u003cb\u003e9\u003c/b\u003e, 27. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/1472-6963-9-27\u003c/span\u003e\u003cspan address=\"10.1186/1472-6963-9-27\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (Feb. 2009).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLarsen, A. T. et al. Estimation of the maximum potential cost saving from reducing serious adverse events in hospitalized patients, \u003cem\u003eActa Anaesthesiol Scand\u003c/em\u003e, vol. 68, no. 10, pp. 1471\u0026ndash;1480, Nov. (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/aas.14525\u003c/span\u003e\u003cspan address=\"10.1111/aas.14525\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoonesinghe, S. R. et al. Survival after postoperative morbidity: a longitudinal observational cohort study. \u003cem\u003eBr. J. Anaesth.\u003c/em\u003e \u003cb\u003e113\u003c/b\u003e (6), 977\u0026ndash;984. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1093/bja/aeu224\u003c/span\u003e\u003cspan address=\"10.1093/bja/aeu224\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (Dec. 2014).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIm, Y. et al. Time-to-antibiotics and clinical outcomes in patients with sepsis and septic shock: a prospective nationwide multicenter cohort study. \u003cem\u003eCrit. Care\u003c/em\u003e. \u003cb\u003e26\u003c/b\u003e (1), 19. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s13054-021-03883-0\u003c/span\u003e\u003cspan address=\"10.1186/s13054-021-03883-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (Dec. 2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNational Early Warning Score (NEWS). 2. Accessed: Apr. 29, 2025. [Online]. Available: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.rcp.ac.uk/improving-care/resources/national-early-warning-score-news-2/\u003c/span\u003e\u003cspan address=\"https://www.rcp.ac.uk/improving-care/resources/national-early-warning-score-news-2/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBriggs, J. et al. Mar., Safer and more efficient vital signs monitoring protocols to identify the deteriorating patients in the general hospital ward: an observational study, \u003cem\u003eHealth Soc Care Deliv Res\u003c/em\u003e, pp. 1\u0026ndash;143, (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3310/HYTR4612\u003c/span\u003e\u003cspan address=\"10.3310/HYTR4612\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eClifton, D. A. et al. Errors\u0026rsquo; and omissions in paper-based early warning scores: the association with changes in vital signs\u0026ndash;a database analysis. \u003cem\u003eBMJ Open.\u003c/em\u003e \u003cb\u003e5\u003c/b\u003e (7), e007376. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1136/bmjopen-2014-007376\u003c/span\u003e\u003cspan address=\"10.1136/bmjopen-2014-007376\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (July 2015).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRowland, B. A., Motamedi, V., Michard, F., Saha, A. K. \u0026amp; Khanna, A. K. Impact of continuous and wireless monitoring of vital signs on clinical outcomes: a propensity-matched observational study of surgical ward patients, \u003cem\u003eBr J Anaesth\u003c/em\u003e, vol. 132, no. 3, pp. 519\u0026ndash;527, Mar. (2024). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.bja.2023.11.040\u003c/span\u003e\u003cspan address=\"10.1016/j.bja.2023.11.040\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDykes, P. C. et al. Reducing ICU Utilization, Length of Stay, and Cost by Optimizing the Clinical Use of Continuous Monitoring System Technology in the Hospital, \u003cem\u003eAm J Med\u003c/em\u003e, vol. 135, no. 3, pp. 337\u0026ndash;341.e1, Mar. (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.amjmed.2021.09.024\u003c/span\u003e\u003cspan address=\"10.1016/j.amjmed.2021.09.024\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEddahchouri, Y. et al. Effect of continuous wireless vital sign monitoring on unplanned ICU admissions and rapid response team calls: a before-and-after study. \u003cem\u003eBr. J. Anaesth.\u003c/em\u003e \u003cb\u003e128\u003c/b\u003e (5), 857\u0026ndash;863. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.bja.2022.01.036\u003c/span\u003e\u003cspan address=\"10.1016/j.bja.2022.01.036\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (May 2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAreia, C. et al. The impact of wearable continuous vital sign monitoring on deterioration detection and clinical outcomes in hospitalised patients: a systematic review and meta-analysis. \u003cem\u003eCrit. Care\u003c/em\u003e. \u003cb\u003e25\u003c/b\u003e (1), 351. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s13054-021-03766-4\u003c/span\u003e\u003cspan address=\"10.1186/s13054-021-03766-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (Sept. 2021).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eM\u0026oslash;lgaard, J. et al. Oct., Continuous Vital Sign Monitoring at the Surgical Ward for Improved Outcomes After Major Noncardiac Surgery: A Randomized Clinical Trial, \u003cem\u003eAnesth Analg\u003c/em\u003e, vol. 141, no. 4, pp. 807\u0026ndash;817, (2025). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1213/ANE.0000000000007606\u003c/span\u003e\u003cspan address=\"10.1213/ANE.0000000000007606\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003e\u0026AElig; et al. Prediction of serious outcomes based on continuous vital sign monitoring of high-risk patients. \u003cem\u003eComput. Biol. Med.\u003c/em\u003e \u003cb\u003e147\u003c/b\u003e, 105559. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.compbiomed.2022.105559\u003c/span\u003e\u003cspan address=\"10.1016/j.compbiomed.2022.105559\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (Aug. 2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHaahr-Raunkjaer, C. et al. Continuous monitoring of vital sign abnormalities; association to clinical complications in 500 postoperative patients. \u003cem\u003eActa Anaesthesiol. Scand.\u003c/em\u003e \u003cb\u003e66\u003c/b\u003e (5), 552\u0026ndash;562. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/aas.14048\u003c/span\u003e\u003cspan address=\"10.1111/aas.14048\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRuppel, H., Dougherty, M., Bonafide, C. P. \u0026amp; Lasater, K. B. Alarm burden and the nursing care environment: a 213-hospital cross-sectional study, \u003cem\u003eBMJ Open Qual\u003c/em\u003e, vol. 12, no. 4, p. e002342, Oct. (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1136/bmjoq-2023-002342\u003c/span\u003e\u003cspan address=\"10.1136/bmjoq-2023-002342\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSendelbach, S. \u0026amp; Funk, M. Alarm fatigue: a patient safety concern. \u003cem\u003eAACN Adv. Crit. Care\u003c/em\u003e, \u003cb\u003e24\u003c/b\u003e, 4, pp. 378\u0026ndash;386 ; quiz 387\u0026ndash;388, 2013, \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1097/NCI.0b013e3182a903f9\u003c/span\u003e\u003cspan address=\"10.1097/NCI.0b013e3182a903f9\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKj\u0026aelig;rgaard, K., M\u0026oslash;lgaard, J., Rasmussen, S. M., Meyhoff, C. S. \u0026amp; Aasvang, E. K. The effect of technical filtering and clinical criteria on alert rates from continuous vital sign monitoring in the general ward, \u003cem\u003eHospital Practice\u003c/em\u003e, vol. 51, no. 5, pp. 295\u0026ndash;302, Oct. (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1080/21548331.2023.2298185\u003c/span\u003e\u003cspan address=\"10.1080/21548331.2023.2298185\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKause, J. et al. A comparison of antecedents to cardiac arrests, deaths and emergency intensive care admissions in Australia and New Zealand, and the United Kingdom\u0026ndash;the ACADEMIA study. \u003cem\u003eResuscitation\u003c/em\u003e \u003cb\u003e62\u003c/b\u003e (3), 275\u0026ndash;282. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.resuscitation.2004.05.016\u003c/span\u003e\u003cspan address=\"10.1016/j.resuscitation.2004.05.016\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (Sept. 2004).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBoulitsakis Logothetis, S., Green, D., Holland, M. \u0026amp; Al Moubayed, N. Predicting acute clinical deterioration with interpretable machine learning to support emergency care decision making. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e13\u003c/b\u003e (1), 13563. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-023-40661-0\u003c/span\u003e\u003cspan address=\"10.1038/s41598-023-40661-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (Aug. 2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAldhoayan, M. D. \u0026amp; Aljubran, Y. Prediction of ICU Patients\u0026rsquo; Deterioration Using Machine Learning Techniques. \u003cem\u003eCureus\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e (5), e38659. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.7759/cureus.38659\u003c/span\u003e\u003cspan address=\"10.7759/cureus.38659\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (May 2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCarlsson, C. J. et al. Apr., Continuous Glucose Monitoring Reveals Perioperative Hypoglycemia in Most Patients With Diabetes Undergoing Major Surgery: A Prospective Cohort Study, \u003cem\u003eAnn Surg\u003c/em\u003e, vol. 277, no. 4, pp. 603\u0026ndash;611, (2023). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1097/SLA.0000000000005246\u003c/span\u003e\u003cspan address=\"10.1097/SLA.0000000000005246\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eElvekjaer, M. et al. Sept., Clinical impact of vital sign abnormalities in patients admitted with acute exacerbation of chronic obstructive pulmonary disease: an observational study using continuous wireless monitoring, \u003cem\u003eIntern Emerg Med\u003c/em\u003e, vol. 17, no. 6, pp. 1689\u0026ndash;1698, (2022). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1007/s11739-022-02988-w\u003c/span\u003e\u003cspan address=\"10.1007/s11739-022-02988-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eM\u0026oslash;lgaard, J. et al. Continuous wireless pre- and postoperative vital sign monitoring reveal new, severe desaturations after vascular surgery. \u003cem\u003eActa Anaesthesiol. Scand.\u003c/em\u003e \u003cb\u003e67\u003c/b\u003e (1), 19\u0026ndash;28. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/aas.14158\u003c/span\u003e\u003cspan address=\"10.1111/aas.14158\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoe, B. P. et al. Boosted decision trees, an alternative to artificial neural networks. \u003cem\u003eNucl. Instrum. Meth A\u003c/em\u003e. \u003cb\u003e543\u003c/b\u003e, 2\u0026ndash;3. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1016/j.nima.2004.12.018\u003c/span\u003e\u003cspan address=\"10.1016/j.nima.2004.12.018\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2005).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen, T. \u0026amp; Guestrin, C. XGBoost: A Scalable Tree Boosting System, in \u003cem\u003eProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining\u003c/em\u003e, Aug. pp. 785\u0026ndash;794. (2016). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1145/2939672.2939785\u003c/span\u003e\u003cspan address=\"10.1145/2939672.2939785\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGr\u0026oslash;nb\u0026aelig;k, K. K. et al. Continuous monitoring is superior to manual measurements in detecting vital sign deviations in patients with COVID-19. \u003cem\u003eActa Anaesthesiol. Scand.\u003c/em\u003e \u003cb\u003e67\u003c/b\u003e (5), 640\u0026ndash;648. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1111/aas.14221\u003c/span\u003e\u003cspan address=\"10.1111/aas.14221\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYoussef Ali Amer, A. et al. Jan., Vital Signs Prediction and Early Warning Score Calculation Based on Continuous Monitoring of Hospitalised Patients Using Wearable Technology, \u003cem\u003eSensors\u003c/em\u003e, vol. 20, no. 22, Art. no. 22, (2020). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3390/s20226593\u003c/span\u003e\u003cspan address=\"10.3390/s20226593\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe, R. \u0026amp; Chiang, J. N. Simultaneous forecasting of vital sign trajectories in the ICU. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cb\u003e15\u003c/b\u003e (1), 14996. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-025-99719-w\u003c/span\u003e\u003cspan address=\"10.1038/s41598-025-99719-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (Apr. 2025).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003e\u003cstrong\u003eTable 1: Patient demography.\u003c/strong\u003e \u003cem\u003eAverage stay duration and demographic information for the patient dataset as well as threshold alert count and most common admission reason.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cimg src=\"https://myfiles.space/user_files/69519_bce2c0439cd956a6/69519_custom_files/img1768847438.png\"\u003e\u003c/em\u003e\u003cbr\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2: p-values for ML models tested against Random and Threshold-based alerts.\u0026nbsp;\u003c/strong\u003e\u003cem\u003eAn overview for the derived p-values from a Fishers Exact test for the BDT model. ML models are all significantly different from a random alert calling system and the threshold-based system.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cimg src=\"https://myfiles.space/user_files/69519_bce2c0439cd956a6/69519_custom_files/img1768847449.png\"\u003e\u003c/em\u003e\u003cbr\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3: Area Under Receiver Operating Curve with 95 % CI, False Positive Rate and True Positive Rates for all methods of alerting and models.\u003cem\u003e\u0026nbsp;\u003c/em\u003e\u003c/strong\u003e\u003cem\u003eOverview showing the calculated AUROC, FPR and TPR for 24-hour and 8-hour models with and without metadata. * Fixed FPR was calculated by using the FPR for the threshold alerts and finding the corresponding [FPR, TPR] coordinate on the AUROC curve for the appropriate model. **Finding the TPR used an identical method to * but using a fixed TPR value instead.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e\u003cimg src=\"https://myfiles.space/user_files/69519_bce2c0439cd956a6/69519_custom_files/img1768847460.png\"\u003e\u003c/em\u003e\u003cbr\u003e\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Machine Learning, Patient Monitoring, Predictive Analysis, Time Series, Clinical Support","lastPublishedDoi":"10.21203/rs.3.rs-8436885/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8436885/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eContinuous Vital Sign Monitoring (CVSM) may allow early detection of patient deterioration with clinical impact from reduced complications. CVSM-based alert algorithms are vulnerable to missing data and artefacts, where large numbers of false positive alerts impose alert fatigue among healthcare staff, especially in low-staffed environments such as general wards. An unexplored option for overcoming missing data and improving alert precision may be using machine learning (ML) models with linear regression to summarize periods of patient data to capture critical vital sign trajectories. Four different ML-analyses of 5 unique vital signs were compared to a state-of-the-art algorithm combining vital sign duration and severity. The data consisted of continuous and semi-continuous vital signs alongside timestamps for physician-curated Serious Adverse Events (SAE) and patient metadata, from 2423 patients monitored during admission for major surgery or acute medical disease. Our results demonstrate that ML-based models improve True Positive Rates (TPR by 0.4) and False Positive Rates (FPR by 0.69). Compared to threshold-based alerts, our approach, showed superior performance in predicting SAE within 24 hours (p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and 8 hours (p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) of an alert. The substantial improvements validate our models\u0026rsquo; competitive edge on clinical data, when directly compared to current alerting systems, without increasing false positives.\u003c/p\u003e","manuscriptTitle":"Machine learning outperforms state-of-the-art continuous vital sign monitoring","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-01-20 08:14:51","doi":"10.21203/rs.3.rs-8436885/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-02-18T14:36:24+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-18T05:22:19+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"161713404794505701569570626422919845283","date":"2026-02-09T08:53:29+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"232309210707116023637890615009825973046","date":"2026-02-06T09:40:00+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-01-29T07:32:38+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"126509255767238613234995550914811196542","date":"2026-01-21T11:44:15+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"206776662146537578738651466057215722468","date":"2026-01-21T04:30:58+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"119498873042574001114899277770532015501","date":"2026-01-16T06:35:20+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"4509222876151991334947493124150615882","date":"2026-01-14T04:03:35+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-01-14T02:30:40+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-01-12T21:38:01+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-01-07T07:26:15+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-01-05T15:05:39+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2026-01-05T14:58:29+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"287646b0-4b53-4234-963c-87b6dc2b616a","owner":[],"postedDate":"January 20th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"in-revision","subjectAreas":[{"id":61223793,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":61223794,"name":"Health sciences/Diseases"},{"id":61223795,"name":"Health sciences/Health care"},{"id":61223796,"name":"Physical sciences/Mathematics and computing"},{"id":61223797,"name":"Health sciences/Medical research"}],"tags":[],"updatedAt":"2026-05-08T06:25:13+00:00","versionOfRecord":[],"versionCreatedAt":"2026-01-20 08:14:51","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8436885","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8436885","identity":"rs-8436885","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}