Abstract
Atrial fibrillation (AF) stands as the predominant arrhythmia observed in ICU patients. Nevertheless, the absence of a swift and precise method for prediction and detection poses a challenge. This study aims to provide a comprehensive literature review on the application of machine learning (ML) algorithms for predicting and detecting new-onset atrial fibrillation (NOAF) in ICU-treated patients. Following the PRISMA recommendations, this systematic review outlines ML models employed in the prediction and detection of NOAF in ICU patients and compares the ML-based approach with clinical-based methods. Inclusion criteria comprised randomized controlled trials (RCTs), observational studies, cohort studies, and case–control studies. A total of five articles published between November 2020 and April 2023 were identified and reviewed to extract the algorithms and performance metrics. Reviewed studies sourced 108,724 ICU admission records form databases, e.g., MIMIC. Eight prediction and detection methods were examined. Notably, CatBoost exhibited superior performance in NOAF prediction, while the support vector machine excelled in NOAF detection. Machine learning algorithms emerge as promising tools for predicting and detecting NOAF in ICU patients. The incorporation of these algorithms in clinical practice has the potential to enhance decision-making and the overall management of NOAF in ICU settings.
Similar content being viewed by others
Avoid common mistakes on your manuscript.
Introduction
Atrial fibrillation (AF) is the most prevalent arrhythmia in the ICU patients [1], with new-onset atrial fibrillation (NOAF) developing in one of six patients admitted to the ICU [2]. Prediction and early detection NOAF are imperative in ICU settings, given its association with life-threatening complications and prolonged hospital length of stay [1]. Specific triggers in ICU patients, such as structural and electrical remodeling related to infection and inflammation, along with arrhythmogenic triggers like continuous catecholamine infusion, can precipitate NOAF [1]. Clinical suspicion and detection of NOAF often relay on continuous ECG monitoring and analysis of 12-leads ECG analysis [2]. Severe acute complications encompass hemodynamic compromise, peripheral organs embolism, ischemia (brain, kidney, etc.), and death [3,4,5]. The occurrence of NOAF is linked to a more challenging clinical course, including extended length of stay increased mortality and the development of ICU-acquired weakness in survivors [5,6,7]. Several models have been proposed to predict and detect NOAF in ICU patients, with the current gold standard being the “post-operative atrial fibrillation (POAF) score’’. However, its clinical utility is constrained by its poor accuracy [8,9,10,11,12].
Machine learning (ML), a subset of artificial intelligence (AI), encompasses algorithms employing statistical and optimization methods to learn from prior data experiences. Its primary goal is to detect or predict valuable outcomes within extensive databases [13,14,15]. The significant ML development is related to the possible replication of human intelligence in machines programmed to duplicate important cognitive processes such as learning, problem-solving, and decision-making to analyze and process large amounts of data, focused on extracting useful information without being explicitly programmed [16]. ML capabilities extend to the development of procedures that autonomously learn from prior experiences, enhancing knowledge in specific domains. ML-based algorithms represent emerging and promising techniques for early AF detection [17]. These algorithms exhibit the capacity to identify patterns, make predictions, and propose actions [18]. Prediction and detection of NOAF are critical in ICU patients holding the potential to substantially contribute to the complicated clinical courses [1].
Despite the expanding body of evidence regarding the clinical applications of ML in risk stratification and early diagnosis of NOAF in ICU patients, there is a noticeable absence of dedicated reports on the role of ML in this specific context. The primary objective of this systematic review (SR) is to elucidate the existing evidence concerning the role of ML in both predicting and detecting NOAF in ICU patients.
Materials and methods
Protocol and registration
This SR was conducted based on the recommendations of the Preferred Reporting and Items for Systematic Reviews and Meta-Analyses (PRISMA) and was recorded in the PROSPERO registry for SR (N. CRD42023397136, Feb 17, 2023).
Search strategy
A literature search was conducted through PubMed, Embase, Scopus, and Medline databases, covering publications available until May 31, 2023. The search strategy employed the following terms: (atrial fibrillation OR new-onset atrial fibrillation) AND (artificial intelligence OR machine learning) AND (intensive care unit OR ICU).
Inclusion criteria comprised randomized controlled trials (RCTs), observational studies, cohort studies, and case–control studies. Full-text articles published in English, focusing on adult patients (age > 18 years old) treated in the ICU and utilizing ML for predicting the clinical outcome of AF, both pre-existing and new onset, were included. Exclusion criteria encompassed case reports, comments, letters to the editor, editorials, study protocols, and replies. Studies involving pediatric patients and those not published in English were also excluded.
The primary outcomes sought were the ML-based prediction and/or detection of NOAF in patients admitted to the ICU. The assessment focused on evaluating the accuracy, sensitivity, and specificity of the machine learning algorithms in predicting and detecting the occurrence of NOAF.
Risk of bias assessment in individual studies
The assessment of bias risk encompassed an examination of five key parameters: a sufficiently sized cohort, appropriate cross-validation, an external validation set, blinding of participants and personnel, and handling of incomplete outcome data (Table 1). For data extraction, a predefined form was utilized, capturing essential information such as study type, sample size, patient characteristics, intervention/exposure details, comparator details, outcome measures, effect measures, follow-up time, funding source, and conflicts of interest.
Results
Study selection and description
Literature search was conducted in July 2023, resulting in the retrieval of 1596 records encompassing risk factors, treatment, prophylaxis, monitoring, and outcomes associated with AF. Among these, 714 articles (44, 7%) were identified as potentially relevant for investigating the role of ML in the prediction and detection of NOAF. Two expert reviewers independently screened the papers, resulting in the selection of 59 articles out of the 714 selected. A subsequent secondary screening focused on excluding studies involving non-ICU patients. Any disparities between the reviewers were resolved through discussion and consensus. Ultimately five studies, each aligning with at least one of the designated outcomes, were deemed eligible for inclusion in the present SR. These studies were further categorized into prediction of NOAF occurrence (n = 2) and detection of NOAF (n = 3) in ICU. The study design is summarized in the PRISMA flowchart reported in Fig. 1.
A comprehensive dataset comprising medical records from a total of 108,724 subjects was sourced from various databases, including Medical Information Mart for Intensive Care (MIMIC)–II [19], –III [20, 21], –IV [22], Kensington General Hospital (KGH) database [23], AmsterdamUMCdb [22], Ghent University Hospital database [22], and Belgian hospital ICU database [22]. The ICU admission records were retrieved for ML models design, testing, and external validation [22, 23]. The records of ICU admissions used in the modeling studies spanned the period from 2001 to 2020 [19,20,21,22,23]. The studies employed eight distinct ML models: K-nearest neighbor (K-NN) and support vector machine (SVM) were utilized in three studies [19,20,21], while forest (RF) and decision tree (DT) were applied in two studies [19, 21]. In addition, logistic reasoning [19], gradient boosted machine [19], CatBoost classification models [22], and deep learning models [23] were each analyzed in single studies.
The results of the selected studies are presented in two distinct sections: ML-based prediction of NOAF and ML detection of NOAF occurrence.
AI-based prediction of NOAF occurrence
The prediction of NOAF occurrence in ICU patients was examined in two retrospective studies [19, 22]. These studies employed database analysis and reported models using the MIMIC-III [16] MIMIC-IV [22], AmsterdamUMC [22], and GUH [22] databases. The patients’ bedside records were collected between 2001 and 2012 [19], 2003 and 2016 [22], 2008–2019 [22] and 2013–2020 [22]. A comprehensive investigation encompassing a total of 107,463 cases is summarized in Table 2. In all the studies, ML-based models were explored, including K-NN, SVM, RF, LR, GBM, and DT [19], with one study incorporating CatBoost classification models [22]. Notably, one of the studies applied ML-based electrocardiogram (ECG) waveform processing techniques [22], while the second study predicted NOAF occurrence based on risk parameters such as age, chronic obstructive pulmonary disease, eGFR of 15 ml/min per 1.73 m2 or dialysis, emergency status, preoperative intra-aortic balloon pump need, valve surgery, and left ventricular ejection fraction < 30% [19].
Each study evaluated the performance of their models using the area under the curve (AUC) (Table 2), achieving the best results of AUC = 0.81 [22] and AUC = 0.74 [19]. The authors of both publications assert the superiority of their models compared to other currently proposed ML-based AF prediction models [19, 22]. In addition, they claim superiority over a current gold standard clinical scoring tool, namely, POAF scoring system, which exhibited an AUC of 0.63 [19].
ML-based detection of NOAF
The detection of NOAF in ICU patients was investigated in three retrospective studies [20, 21, 23] based on database analysis report models: MIMIC-III [20, 21]. The medical records were collected between 2001–2012 [20] and 2015–2020 [23]. A total of 1261 cases were investigated (Table 3). The three studies investigated ML-based models: K-NN [20, 21], SVM [20, 21], RF [20], and DA [21]. One study [23] did not explicitly specify the type of ML-based method utilized. Notably, all studies incorporated ML-based electrocardiogram (ECG) waveform processing techniques [20, 21, 23]. The performance evaluation in each study was conducted using various metrics. The best values, as presented in Table 3, were achieved by the study using F1 score (0.64–0.67) and expected calibration error (0.05–0.07) [23]. In another study, the SVM classifier demonstrated sensitivity greater than 98% and specificity exceeding 93% [20]. The authors of the three publications claim that their models are superior compared to other currently proposed ML-based AF detection models pointing out at a better sensitivity [23] and accuracy [20, 21].
Discussion
This SR originally reports the available evidence on the role of ML-based prediction and detection models in patients with NOAF treated in ICU.
The first study compares ML-based predictive models capable of anticipating NOAF based on an extensive set of 194 variables including administered drug therapy, laboratory values, anthropometric data, and hemodynamic parameters. Notably, the designed models demonstrate the ability to predict NOAF onset 24–36 h before its occurrence [22]. The second study presents ML-based models that rely their prediction on demographics, physiological parameters, laboratory results, and clinical outcomes [19]. Although the specific timeframe for prediction is not explicitly specified in this article, the study remains valuable in the ICU clinical practice as it provides crucial information to intensivists regarding patients at risk of developing complications, such as NOAF. Based on the comparison of the performance of each ML-model measured in terms of AUC, it is observed that all ML models outperform traditional clinical scores and among the studied algorithms, CatBoost shows to be superior in comparison to the others. A noteworthy observation from the comparison of the ML models’ performance, measured in terms of the area under the curve (AUC), indicates that all ML models surpass traditional clinical scores. Furthermore, among the studied algorithms, CatBoost emerges as superior in comparison to the others. The selected articles suggest the potential capability of the described ML models to predict and detect NOAF, signaling a promising potential for a possible implementation of ML in both research and clinical practice. Importantly, ML methods exhibit an advantage over conventional clinical scoring systems, particularly in predicting the occurrence of NOAF and significantly limits human error and shortens detection process in unclear ECG readings. Moreover, ML methodologies significantly mitigate human error and expedite the detection process, especially in cases of unclear ECG readings.
Numerous research findings, comprehensive meta-analyses, and systematic reviews, consistently affirm the substantial potential in the utilization of ML methods. A study conducted in non-ICU setting stands as a pertinent example of leveraging ML for predicting NOAF [24]. Reliable outcomes have consistently demonstrated the effectiveness of ML predicting models, particularly in their ability to anticipate intensive care delirium up to 12 h in advance [25]. In addition, the use of ML has shown promise in predicting cardiovascular complications in diabetic patients [26].
Time factor is crucial in diagnosing NOAF in critical patients, given the elevated risk of death and the potential long-term complications [1,2,3,4,5,6,7]. The prediction and rapid detection of NOAF have been highly discussed in the medical society. The medical community has extensively discussed the importance of predicting and rapidly detecting NOAF. Numerous alternative strategies have been proposed for NOAF prediction, including assessing DNA methylation levels [27], measuring plasma aldosterone concentrations [28], and exploring immune-associated biomarkers [29]. While these methods indicate potential for NOAF prediction, further investigations through prospective clinical trials are essential. In this context, rapidly developing ML-based methods and their increasing integration into medical practice demonstrate a remarkable potential for the prediction and detection of NOAF. The agility and adaptability of ML approaches offer a promising avenue for timely identification and management of NOAF, potentially improving patient outcomes in critical care settings. Therefore, the evolving ML-based methods and their growing integration into medical practice hold substantial potential for the prediction and detection of NOAF.
Limitations of this SR include the relatively limited number of available studies, all of which were retrospective and based on few databases which could result in a partially overlapping of records. To address this, future prospective randomized clinical trials comparing the performance of ML-based models versus traditional clinical methods for the prediction and detection of NOAF in ICU are necessary. Such trials should provide more robust evidence and contribute to the refinement of ML applications in critical care settings.
Conclusion
This SR provides a comprehensive summary of all available evidence related to the prediction and detection of NOAF in the ICU. Importantly, the evidence suggests that ML-based methods have already surpassed the POAF score, which currently serves as the gold standard in clinical practice. Among the various ML algorithms studied, CatBoost emerges as the top performer. Furthermore, the rapid development of ML signifies the potential for a paradigm shift, advocating for the redesign or integration of ML-based methods either independently or in conjunction with traditional risk scoring systems in the ICU. This recommendation is underscored by the need for further evaluation in prospective, randomized controlled trials, which will provide more robust insights into the efficacy and feasibility of ML applications in the critical care setting.
Data availability
All data associated with this manuscript are included in the main text and Supplementary Materials.
Abbreviations
- AF:
-
Atrial fibrillation
- NOAF:
-
New-onset atrial fibrillation
- ICU:
-
Intensive care unit
- ML:
-
Machine learning
- RF:
-
Random forest
- DT:
-
Decision tree
- LR:
-
Logistic regression
- K-NN:
-
K-nearest neighbors
- SVM:
-
Support vector machine
- GBM:
-
Gradient boosted machine
- AUC:
-
Area under the curve
- ECE:
-
Expected calibration error
- ESCE:
-
Expected signed calibration error
- PRIMSA:
-
Preferred Reporting and Items for Systematic Reviews and Meta-Analyses
- SR:
-
Systematic review
References
Bosch NA, Cimini J, Walkey AJ. Atrial fibrillation in the ICU. Chest. 2018;154(6):1424–34.
Wetterslev M, Hylander Møller M, Granholm A, Hassager C, Haase N, Lange T, Myatra SN, Hästbacka J, Arabi YM, Shen J, Cronhjort M, Lindqvist E, Aneman A, Young PJ, Szczeklik W, Siegemund M, Koster T, Aslam TN, Bestle MH, Girkov MS, Kalvit K, Mohanty R, Mascarenhas J, Pattnaik M, Vergis S, Haranath SP, Shah M, Joshi Z, Wilkman E, Reinikainen M, Lehto P, Jalkanen V, Pulkkinen A, An Y, Wang G, Huang L, Huang B, Liu W, Gao H, Dou L, Li S, Yang W, Tegnell E, Knight A, Czuczwar M, Czarnik T, Perner A, AFIB-ICU Collaborators. Atrial fibrillation (AFIB) in the ICU: incidence, risk factors, and outcomes: the international AFIB-ICU cohort study. Crit Care Med. 2023;51(9):1124–37.
Santhanakrishnan R, Wang N, Larson MG, Magnani JW, McManus DD, Lubitz SA, Ellinor PT, Cheng S, Vasan RS, Lee DS, Wang TJ, Levy D, Benjamin EJ, Ho JE. Atrial fibrillation begets heart failure and vice versa: temporal associations and differences in preserved versus reduced ejection fraction. Circulation. 2016;133(5):484–92.
Benjamin EJ, Wolf PA, D’Agostino RB, Silbershatz H, Kannel WB, Levy D. Impact of atrial fibrillation on the risk of death: the Framingham Heart Study. Circulation. 1998;98(10):946–52.
Klein Klouwenberg PM, Frencken JF, Kuipers S, Ong DS, Peelen LM, van Vught LA, Schultz MJ, van der Poll T, Bonten MJ, Cremer OL, MARS Consortium. Incidence, predictors, and outcomes of new-onset atrial fibrillation in critically ill patients with sepsis. A cohort study. Am J Respir Crit Care Med. 2017;195(2):205–11.
Moss TJ, Calland JF, Enfield KB, Gomez-Manjarres DC, Ruminski C, DiMarco JP, Lake DE, Moorman JR. New-onset atrial fibrillation in the critically ill. Crit Care Med. 2017;45(5):790–7.
Jolley SE, Bunnell AE, Hough CL. ICU-acquired weakness. Chest. 2016;150(5):1129–40.
Mariscalco G, Engström KG. Atrial fibrillation after cardiac surgery: risk factors and their temporal relationship in prophylactic drug strategy decision. Int J Cardiol. 2008;129(3):354–62.
Amar D, Shi W, Hogue CW Jr, Zhang H, Passman RS, Thomas B, Bach PB, Damiano R, Thaler HT. Clinical prediction rule for atrial fibrillation after coronary artery bypass grafting. J Am Coll Cardiol. 2004;44(6):1248–53.
Mathew JP, Fontes ML, Tudor IC, Ramsay J, Duke P, Mazer CD, Barash PG, Hsu PH, Mangano DT, Investigators of the Ischemia Research and Education Foundation, Multicenter Study of Perioperative Ischemia Research Group. Mangano DT A multicenter risk index for atrial fibrillation after cardiac surgery. JAMA. 2004;291(14):1720–9.
Thorén E, Hellgren L, Jidéus L, Ståhle E. Prediction of postoperative atrial fibrillation in a large coronary artery bypass grafting cohort. Interact Cardiovasc Thorac Surg. 2012;14(5):588–93.
Mariscalco G, Biancari F, Zanobini M, Cottini M, Piffaretti G, Saccocci M, Banach M, Beghi C, Angelini GD. Bedside tool for predicting the risk of postoperative atrial fibrillation after cardiac surgery: the POAF score. J Am Heart Assoc. 2014;3(2): e000752.
Viderman D, Abdildin YG, Batkuldinova K, Badenes R, Bilotta F. Artificial intelligence in resuscitation: a scoping review. J Clin Med. 2023;12(6):2254. https://doi.org/10.3390/jcm12062254. (PMID:36983255;PMCID:PMC10054374).
Haug CJ, Drazen JM. Artificial intelligence and machine learning in clinical medicine, 2023. N Engl J Med. 2023;388(13):1201–8. https://doi.org/10.1056/NEJMra2302038. (PMID: 36988595).
Bellini V, Valente M, Bertorelli G, Pifferi B, Craca M, Mordonini M, Lombardo G, Bottani E, Del Rio P, Bignami E. Machine learning in perioperative medicine: a systematic review. J Anesth Analg Crit Care. 2022;2(1):2. https://doi.org/10.1186/s44158-022-00033-y. (PMID:37386544;PMCID:PMC8761048).
Rajkomar A, Dean J, Kohane I. Machine learning in medicine. N Engl J Med. 2019;380(14):1347–58.
Harmon DM, Sehrawat O, Maanja M, Wight J, Noseworthy PA. Artificial intelligence for the detection and treatment of atrial fibrillation. Arrhythm Electrophysiol Rev. 2023;12: e12.
Jentzer JC, Kashou AH, Murphree DH. Clinical applications of artificial intelligence and machine learning in the modern cardiac intensive care unit. Intell Based Med. 2023;7:100089.
Karri R, Kawai A, Thong YJ, Ramson DM, Perry LA, Segal R, Smith JA, Penny-Dimri JC. Machine learning outperforms existing clinical scoring tools in the prediction of postoperative atrial fibrillation during intensive care unit admission after cardiac surgery. Heart Lung Circ. 2021;30(12):1929–37.
Bashar SK, Han D, Zieneddin F, Ding E, Fitzgibbons TP, Walkey AJ, McManus DD, Javidi B, Chon KH. Novel density poincaré plot based machine learning method to detect atrial fibrillation from premature atrial/ventricular contractions. IEEE Trans Biomed Eng. 2021;68(2):448–60.
Bashar SK, Hossain MB, Ding E, Walkey AJ, McManus DD, Chon KH. Atrial fibrillation detection during sepsis: study on MIMIC III ICU data. IEEE J Biomed Health Inform. 2020;24(11):3124–35.
Verhaeghe J, De Corte T, Sauer CM, Hendriks T, Thijssens OWM, Ongenae F, Elbers P, De Waele J, Van Hoecke S. Generalizable calibrated machine learning models for real-time atrial fibrillation risk prediction in ICU patients. Int J Med Inform. 2023;175:105086.
Chen B, Javadi G, Hamilton A, Sibley S, Laird P, Abolmaesumi P, Maslove D, Mousavi P. Quantifying deep neural network uncertainty for atrial fibrillation detection with limited labels. Sci Rep. 2022;12(1):20140.
Gue Y, Correa E, Thompson JLP, Homma S, Qian M, Lip GYH. Machine learning predicting atrial fibrillation as an adverse event in the Warfarin and aspirin in reduced cardiac ejection fraction (WARCEF) Trial. Am J Med. 2023;21:S0002-9343.
Gong KD, Lu R, Bergamaschi TS, Sanyal A, Guo J, Kim HB, Nguyen HT, Greenstein JL, Winslow RL, Stevens RD. Predicting intensive care delirium with machine learning: model development and external validation. Anesthesiology. 2023;138(3):299–311.
N, Abdul Murad NA, Chin SF, Jaafar R, Abdullah N. Cardiovascular complications in a diabetes prediction model using machine learning: a systematic review. Cardiovasc Diabetol. 2023, 22(1):13.
Fischer MA, Mahajan A, Cabaj M, Kimball TH, Morselli M, Soehalim E, Chapski DJ, Montoya D, Farrell CP, Scovotti J, Bueno CT, Mimila NA, Shemin RJ, Elashoff D, Pellegrini M, Monte E, Vondriska TM. DNA methylation-based prediction of post-operative atrial fibrillation. Front Cardiovasc Med. 2022;9:837725.
Chequel M, Ollitrault P, Saloux E, Parienti JJ, Fischer MO, Desgué J, Allouche S, Milliez P, Alexandre J. Preoperative plasma aldosterone levels and postoperative atrial fibrillation occurrence following cardiac surgery: a review of literature and design of the ALDO-POAF study (ALDOsterone for prediction of post-operative atrial fibrillation). Curr Clin Pharmacol. 2016;11(3):150–8.
Zhou Y, Wu Q, Ni G, Hong Y, Xiao S, Liu C, Yu Z. Immune-associated pivotal biomarkers identification and competing endogenous RNA network construction in post-operative atrial fibrillation by comprehensive bioinformatics and machine learning strategies. Front Immunol. 2022;13:974935.
Liberati A, Altman DG, Tetzlaff J, Mulrow C, Gøtzsche PC, Ioannidis JP, Clarke M, Devereaux PJ, Kleijnen J, Moher D. The PRISMA statement for reporting systematic reviews and meta-analyses of studies that evaluate healthcare interventions: explanation and elaboration. BMJ. 2009;339:b2700. https://doi.org/10.1136/bmj.b2700. (PMID:19622552;PMCID:PMC2714672).
Funding
Open access funding provided by Università degli Studi di Roma La Sapienza within the CRUI-CARE Agreement.
Author information
Authors and Affiliations
Contributions
KG and FB conceived of the study idea. FB coordinated the systematic review. KG and FB designed the search strategy. KG and JDS screened abstracts and full texts. KG and JDS acquired the data and judged risk of bias in the studies. KG and LM verified the data and performed the analyses. All authors interpreted the data analyses. All authors cowrote and revised the manuscript for intellectual content. All authors provided their final approval for manuscript submission. All authors agree to be accountable for all aspects of the work.
Corresponding author
Ethics declarations
Conflict of interest
The authors declare that they have no competing interests.
Ethics approval and consent to participate
Not applicable.
Consent for publication
All authors have read the manuscript and consented for this manuscript to be published by Journal of Anesthesia.
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Supplementary Information
Below is the link to the electronic supplementary material.
Rights and permissions
This article is published under an open access license. Please check the 'Copyright Information' section either on this page or in the PDF for details of this license and what re-use is permitted. If your intended use exceeds what is permitted by the license or if you are unable to locate the licence and re-use information, please contact the Rights and Permissions team.
About this article
Cite this article
Glaser, K., Marino, L., Stubnya, J.D. et al. Machine learning in the prediction and detection of new-onset atrial fibrillation in ICU: a systematic review. J Anesth 38, 301–308 (2024). https://doi.org/10.1007/s00540-024-03316-6
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00540-024-03316-6