Abstract:Biomarkers in MedicineVol. 13, No. 15 EditorialFree AccessMachine learning method for the management of acute kidney injury: more than just treating biomarkers individuallyZhongheng ZhangZhongheng Zhang *Author for correspondence: E-mail Address: zh_zhang1984@zju.edu.cnhttps://orcid.org/0000-0002-2336-5323Department of Emergency Medicine, Sir Run Run Shaw Hospital, Zhejiang University School of Medicine, Hangzhou, 310016, PR ChinaPublished Online:27 Sep 2019https://doi.org/10.2217/bmm-2019-0363AboutSectionsView ArticleSupplemental MaterialPDF/EPUB ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareShare onFacebookTwitterLinkedInReddit View articleKeywords: acute kidney injurybig databiomarkerscritical caremachine learningpredictionAcute kidney injury (AKI) is frequently encountered in the intensive care unit (ICU). The occurrence of AKI is of clinical importance because there is strong evidence that even a mild increase in serum creatinine is associated with significantly increased risk of mortality [1]. The clinical definition of AKI is largely dependent on serum creatinine and urine output. However, the increase in serum creatinine usually indicates that the renal function has been severely impaired, and interventions to protect renal functions are usually delayed based on this definition. Moreover, the serum levels of creatinine can be influenced by multiple factors other than renal function. Thus, strenuous efforts have been made to predict AKI accurately and timely in high-risk patients. Many novel biomarkers were proposed to substitute creatinine in the evaluation of renal function, such as cystatin C [2], NGAL, TIM-2, KIM-1 and IGFBP-7 [3]. Most of these biomarkers were investigated independently, and their diagnostic performances were found to be moderate to high (area under the curve: 0.87; 95% CI: 0.81–0.93). However, an individual biomarker is limited in its prediction for AKI, in other words, the development of AKI is multifactorial. I propose that the predictive performance can be further improved with the use of advanced machine learning (ML) techniques.Recent years have witnessed rapid advances in the ML techniques. Generally speaking, the ML approach comprises three broad categories according to their purposes. Supervised learning refers to the algorithm to train a function mapping from feature space to the label (outcome). All observations (subjects) should be labeled in advance. Some sophisticated supervised ML methods include neural networks, decision trees, random forest and regression models. The task to develop a model for the prediction of AKI also belongs to the supervised ML. The advantage of ML such as neural network and decision trees is that they can automatically handle the nonlinear terms, as well as interactions among features. In the era of big data, tens of thousands of features of a subject can be collected and the complex relationships cannot be fully captured with the conventional regression model. The disadvantage of these ML methods are that the underlying relationship between covariates and outcome cannot be fully understand by human beings [4]. Conventionally, generalized linear model is employed to develop risk prediction models in clinical medicine. The advantage of such kind of models is that the coefficient of each covariate can be intuitively interpreted because the exponentiation of the coefficient gives the odds ratio (eβ = odds ratio). In this vein, the models also help to understand the causal relationship between covariates and outcome. The ML methods are widely used for the prediction of AKI in critical care setting. Supplementary Table 1 lists studies employing ML to predict AKI published in 2019 [5–11]. In these studies, various ML methods such as gradient boosted tree, k-nearest neighbors and Naive Bayes were used to predict AKI. The reported accuracies and area under the curves were quite different among the studies, which was attributable to the differences in the study population, the predictors, validation methods and ML algorithms. A systematic review and meta-analysis can be helpful in exploring the potential heterogeneity of the high-risk patients.The second method is the unsupervised ML, which are typically trained in unlabeled observations. Intuitively, unsupervised ML aims to identify latent subclasses by features of the subjects, and there is no supervisor to guide the training iteration. Unsupervised ML encompasses principal component analysis, clustering analysis, latent profile (class) analysis and latent Markov models [12–14]. The unsupervised ML plays an important role in the exploration of the heterogeneity in clinical researches. Although the randomized controlled trial only reported an average effect in the study population, the latent class analysis can help to discover underlying latent subgroups that respond differently to specific interventions. For example, Bhatraju et al. identified two molecularly distinct AKI subphenotypes with different clinical outcomes and responses to vasopressin therapy [15]. Such subphenotypes have important clinical relevance because different subphenotypes respond differently to interventions. Moreover, clinical trials testing different drugs can be performed in different subphenotypes.The third category of ML is the reinforcement learning (RL) method, in which the agent receives a reward after taking an action in a certain state. Thus, the value of a reward is a function of the action and state space. The aim of the agent is to maximize the long-term reward, also known as the return, by responding properly to the state space. RL has been successfully used to individualize treatment strategy for the management of sepsis and mechanical ventilation [16,17]. However, to the best of my knowledge, RL has never been used in the management of AKI. RL is a robust algorithm that can help to form a heuristic approach in the management of AKI. In the design of an AKI study with RL method, the state space can be constructed by using demographics, laboratory variables, hemodynamics and vital signs, and the action space can be formed by vasopressors, fluid strategy and diuretics [18]. In this framework, the heuristic treatment strategy is explored as a whole, which is different from the conventional approach that each specific intervention is explored individually.Due to the rapid development of the ML technique in computer science, it will certainly advance the clinical practice to facilitate patient management. Many modern ICUs have equipped with the early warning system that can help clinicians to identify high-risk patients. However, many of these early warning systems are based on knowledge and fixed rules, but not based on artificial intelligence. In the future, the prediction model established by sophisticated ML algorithms can be more readily to adapt to a certain working environment (i.e., it can learn from the big data provided from electronic healthcare records) and more accurately predict clinical outcomes. Furthermore, these algorithms can also help to recommend individualized treatment strategy if the RL algorithm is deployed in the ICU.In conclusion, the ML approach provides a large number of mathematical algorithms to help to establish and realize precise medicine in the management of AKI. Clinical investigators should familiarize themselves to the rapidly developing field so that best ML model can be tailored to a specific clinical scenario.Supplementary dataTo view the supplementary data that accompany this paper please visit the journal website at: www.futuremedicine.com/doi/suppl/10.2217/bmm-2019-0363Financial & competing interests disclosureZ Zhang received funding from Zhejiang Engineering Research Center of Intelligent Medicine (grant number 2016E10011) from the First Affiliated Hospital of Wenzhou Medical University, The public welfare research project of Zhejiang province (grant number LGF18H150005), the National Natural Science Foundation of China (grant number 81901929) and Scientific research project of Zhejiang Education Commission (grant number Y201737841). The author has no other relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed.No writing assistance was utilized in the production of this manuscript.References1. Kidney Disease Improving Global Outcomes Acute Kidney Injury Work Group. KDIGO clinical practice guideline for acute kidney injury. Kidney Int. 2(Suppl. 1), 1–138 (2012).Google Scholar2. Zhang Z, Lu B, Sheng X, Jin N. Cystatin C in prediction of acute kidney injury: a systemic review and meta-analysis. Am. J. Kidney Dis. 58(3), 356–365 (2011).Crossref, Medline, CAS, Google Scholar3. Beker BM, Corleto MG, Fieiras C, Musso CG. Novel acute kidney injury biomarkers: their characteristics, utility and concerns. Int. Urol. Nephrol. 50(4), 705–713 (2018).Crossref, Medline, CAS, Google Scholar4. Zhang Z, Beck MW, Winkler DA et al. Opening the black box of neural networks: methods for interpreting neural network models in clinical applications. Ann. Transl. Med. 6(11), 216–216 (2018).Crossref, Medline, Google Scholar5. Parreco J, Soe-Lin H, Parks JJ et al. Comparing machine learning algorithms for predicting acute kidney injury. Am. Surg. 85(7), 725–729 (2019).Crossref, Medline, Google Scholar6. Tran NK, Sen S, Palmieri TL et al. Artificial intelligence and machine learning for predicting acute kidney injury in severely burned patients: a proof of concept. Burns 45(6), 1350– 1358 ( 2019).Crossref, Medline, Google Scholar7. He J, Hu Y, Zhang X, Wu L, Waitman LR, Liu M. Multi-perspective predictive modeling for acute kidney injury in general hospital populations using electronic medical records. JAMIA Open 2(1), 115–122 (2019).Crossref, Medline, Google Scholar8. Adhikari L, Ozrazgat-Baslanti T, Ruppert M et al. Improved predictive models for acute kidney injury with IDEA: Intraoperative Data Embedded Analytics. PLoS ONE 14(4), e0214904 (2019).Crossref, Medline, CAS, Google Scholar9. Zimmerman LP, Reyfman PA, Smith ADR et al. Early prediction of acute kidney injury following ICU admission using a multivariate panel of physiological measurements. BMC Med. Inform. Decis. Mak. 19(Suppl. 1), 16–12 (2019).Crossref, Medline, Google Scholar10. Ibrahim NE, McCarthy CP, Shrestha S et al. A clinical, proteomics, and artificial intelligence-driven model to predict acute kidney injury in patients undergoing coronary angiography. Clin. Cardiol. 42(2), 292–298 (2019).Crossref, Medline, Google Scholar11. Tomašev N, Glorot X, Rae JW et al. A clinically applicable approach to continuous prediction of future acute kidney injury. Nature 572(7767), 116–119 (2019).Crossref, Medline, CAS, Google Scholar12. Collins LM, Lanza ST. Latent class and latent transition analysis: With applications in the social, behavioral, and health sciences. John Wiley & Sons, Inc., NJ, USA (2009).Crossref, Google Scholar13. Zhang Z, Abarda A, Contractor AA, Wang J, Dayton CM. Exploring heterogeneity in clinical trials with latent class analysis. Ann. Transl. Med. 6(7), 119–119 (2018).Crossref, Medline, Google Scholar14. Pennoni F, Barbato M, Del Zoppo S. A latent Markov model with covariates to study unobserved heterogeneity among fertility patterns of couples employing natural family planning methods. Front. Public Health 5, 186 (2017).Crossref, Medline, Google Scholar15. Bhatraju PK, Zelnick LR, Herting J et al. Identification of acute kidney injury subphenotypes with differing molecular signatures and responses to vasopressin therapy. Am. J. Respir. Crit. Care Med. 199(7), 863–872 (2019).Crossref, Medline, CAS, Google Scholar16. Komorowski M, Celi LA, Badawi O, Gordon AC, Faisal AA. The Artificial Intelligence Clinician learns optimal treatment strategies for sepsis in intensive care. Nat. Med. 24(11), 1716–1720 (2018).Crossref, Medline, CAS, Google Scholar17. Prasad N, Cheng L-F, Chivers C, Draugelis M, Engelhardt BE. A reinforcement learning approach to weaning of mechanical ventilation in intensive care units. ArXiv. 19(Suppl. 2), 57 (2017).Google Scholar18. Silver D, Huang A, Maddison CJ et al. Mastering the game of Go with deep neural networks and tree search. Nature 529(7587), 484–489 (2016).Crossref, Medline, CAS, Google ScholarFiguresReferencesRelatedDetailsCited ByArtificial Intelligence and Machine Learning in Perioperative Acute Kidney InjuryAdvances in Kidney Disease and Health, Vol. 30, No. 1Application of Machine Learning Algorithms to Predict Acute Kidney Injury in Elderly Orthopedic Postoperative Patients1 March 2022 | Clinical Interventions in Aging, Vol. Volume 17Urinary Metabolome Analyses of Patients with Acute Kidney Injury Using Capillary Electrophoresis-Mass Spectrometry30 September 2021 | Metabolites, Vol. 11, No. 10Blockchain Based Peer-Review Interfaces for Digital Medicine5 March 2020 | Frontiers in Blockchain, Vol. 3 Vol. 13, No. 15 Follow us on social media for the latest updates Supplemental MaterialsMetrics History Received 18 August 2019 Accepted 23 August 2019 Published online 27 September 2019 Published in print October 2019 Information© 2019 Future Medicine LtdKeywordsacute kidney injurybig databiomarkerscritical caremachine learningpredictionSupplementary dataTo view the supplementary data that accompany this paper please visit the journal website at: www.futuremedicine.com/doi/suppl/10.2217/bmm-2019-0363Financial & competing interests disclosureZ Zhang received funding from Zhejiang Engineering Research Center of Intelligent Medicine (grant number 2016E10011) from the First Affiliated Hospital of Wenzhou Medical University, The public welfare research project of Zhejiang province (grant number LGF18H150005), the National Natural Science Foundation of China (grant number 81901929) and Scientific research project of Zhejiang Education Commission (grant number Y201737841). The author has no other relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript apart from those disclosed.No writing assistance was utilized in the production of this manuscript.PDF download

Abstract:The remarkable evolution of surgical techniques has broadened the scope of treatable conditions. From interventional procedures to open surgeries, the surgical landscape is increasingly complex. These advances, although offering new hope for patients, also amplify the challenges clinicians face in providing optimal perioperative care, given the diverse nature of surgeries and the varying health statuses of patients. The Article by David Ouyang and colleagues' published in this issue of The Lancet Digital Health1Ouyang D Theurer J Stein NR et al.Electrocardiographic deep learning for predicting post-procedural mortality: a model development and validation study.Lancet Digit Health. 2023; (published online Dec 7.)https://doi.org/10.1016/S2589-7500(23)00220-0Summary Full Text Full Text PDF PubMed Scopus (0) Google Scholar shows that a deep-learning algorithm interpreting preoperative electrocardiograms (ECGs) can improve discrimination of postoperative mortality. The model was trained and internally validated on a cohort of 45 969 preoperative patients at Cedars-Sinai Medical Center (with an area under the receiver operating characteristic curve [AUC] of 0·83) and further externally validated on two separate health-care system cohorts (the Stanford Healthcare cohort [AUC 0·79] and the Columbia University Medical Center cohort [AUC 0·75]), without any additional further fine-tuning or training. Preoperative risk stratification serves as the cornerstone for perioperative management. Early identification of high-risk patients allows health-care professionals to comprehensively assess the risks and benefits of surgery and devise individualised treatment strategies. Although numerous risk assessment tools have been established, they come with evident limitations. One major drawback of traditional tools is their procedure-specific nature, resulting in the need for distinct models to assess different surgical procedures. For instance, general surgeries use the P-POSSUM score, whereas cardiovascular surgeries require the EuroSCORE or STS scoring systems.2Prytherch DR Whiteley MS Higgins B Weaver PC Prout WG Powell SJ POSSUM and Portsmouth POSSUM for predicting mortality. Physiological and Operative Severity Score for the enUmeration of Mortality and morbidity.Br J Surg. 1998; 85: 1217-1220Crossref PubMed Scopus (558) Google Scholar, 3Ad N Holmes SD Patel J Pritchard G Shuman DJ Halpin L Comparison of EuroSCORE II, original EuroSCORE, and the Society of Thoracic Surgeons risk score in cardiac surgery patients.Ann Thorac Surg. 2016; 102: 573-579Summary Full Text Full Text PDF PubMed Scopus (157) Google Scholar Traditional scoring systems use numerous variables, such as medical history, biomarkers (eg, creatinine, cardiac myosin, and brain natriuretic peptide), ECGs, and imaging data, with some requiring postoperative details for risk adjustment. The complexity of these models contributes to an increased financial burden on patients and increased workload on clinicians. However, most traditional scoring systems exhibit only mediocre predictive accuracy,4Kivrak S Haller G Scores for preoperative risk evaluation of postoperative mortality.Best Pract Res Clin Anaesthesiol. 2021; 35: 115-134Crossref PubMed Scopus (16) Google Scholar causing challenges in meeting the actual requirements of clinical practice. ECG provides vital cardiac information and is easily accessible in clinics. Doctors diagnose heart conditions by analysing ECG waveforms. Advancements in deep learning and abundant ECG data enable computers to detect subtle changes imperceptible to humans, rejuvenating the practice of ECG. Nowadays, the application of artificial intelligence (AI) in ECG is becoming increasingly widespread.5Siontis KC Noseworthy PA Attia ZI Friedman PA Artificial intelligence-enhanced electrocardiography in cardiovascular disease management.Nat Rev Cardiol. 2021; 18: 465-478Crossref PubMed Scopus (240) Google Scholar It improves early prediction of cardiovascular diseases (eg, arrhythmia6Noseworthy PA Attia ZI Behnken EM et al.Artificial intelligence-guided screening for atrial fibrillation using electrocardiogram during sinus rhythm: a prospective non-randomised interventional trial.Lancet. 2022; 400: 1206-1212Summary Full Text Full Text PDF PubMed Scopus (34) Google Scholar and coronary heart disease) and heart structural issues and non-cardiovascular conditions (eg, electrolyte imbalance and anaemia), alongside predicting demographic factors (eg, age and sex). The advancement in the application of AI greatly improves disease screening and daily health management. Ouyang and colleagues' work further expands the boundaries of ECGs' ability. They built a deep-learning model named PreOpNet to predict post-procedural mortality from preoperative ECG signals.1Ouyang D Theurer J Stein NR et al.Electrocardiographic deep learning for predicting post-procedural mortality: a model development and validation study.Lancet Digit Health. 2023; (published online Dec 7.)https://doi.org/10.1016/S2589-7500(23)00220-0Summary Full Text Full Text PDF PubMed Scopus (0) Google Scholar The PreOpNet model stands out for its simplicity and effectiveness in preoperative risk assessment. Using routine preoperative ECGs, the model achieves high accuracy in predicting postoperative mortality and cardiovascular events, and its accuracy remains robust across cardiac surgeries, non-cardiac surgeries, and interventions in the catheterisation laboratory or endoscopy suite.1Ouyang D Theurer J Stein NR et al.Electrocardiographic deep learning for predicting post-procedural mortality: a model development and validation study.Lancet Digit Health. 2023; (published online Dec 7.)https://doi.org/10.1016/S2589-7500(23)00220-0Summary Full Text Full Text PDF PubMed Scopus (0) Google Scholar With ECG-based automated risk assessment, PreOpNet not only simplifies the preoperative evaluation process but also reduces the need for extensive tests and examinations. Consequently, patients will experience a quicker preoperative assessment, with shorter hospitalisation and lower expense, which might translate into considerable health economic value.7Meara JG Leather AJM Hagander L et al.Global Surgery 2030: evidence and solutions for achieving health, welfare, and economic development.Lancet. 2015; 386: 569-624Summary Full Text Full Text PDF PubMed Scopus (2181) Google Scholar Moreover, PreOpNet's lightweight architecture is a pivotal feature that enhances its accessibility and applicability. PreOpNet can stably run on standard computers without the need for specialised or high-performance computing resources. This feature ensures its potential application in underdeveloped regions with scarce access to advanced technological infrastructure, bridging the gap in preoperative risk assessment across different levels of hospitals. However, it is crucial to note that in underdeveloped regions, surgeries mainly consist of low-risk procedures due to the establishment of referral systems.7Meara JG Leather AJM Hagander L et al.Global Surgery 2030: evidence and solutions for achieving health, welfare, and economic development.Lancet. 2015; 386: 569-624Summary Full Text Full Text PDF PubMed Scopus (2181) Google Scholar Because of the scarcity of relevant data, PreOpNet might face challenges in accurately assessing low-risk patients. Therefore, it is necessary to conduct targeted fine-tuning and validation for low-risk patients in future studies, which might further enhance its applicability and practicality. The application of AI in ECG analysis will bring many opportunities to the health-care system. First, clinical researchers need to catch up to embrace this technology quickly. Enhancing training and education would ensure that medical professionals thoroughly understand and correctly use this technology. Second, integrating AI ECG technology with hospital information systems or wearable devices8Attia ZI Harmon DM Dugan J et al.Prospective evaluation of smartwatch-enabled detection of left ventricular dysfunction.Nat Med. 2022; 28: 2497-2503Crossref PubMed Scopus (21) Google Scholar enables real-time ECG analysis and continuous learning and monitoring.9Sun C Li H Song M Cai D Zhang B Hong S Continuous diagnosis and prognosis by controlling the update process of deep neural networks.Patterns (N Y). 2023; 4100687Google Scholar Third, government agencies also need to establish appropriate standards for evaluating AI medical devices.10Wu E Wu K Daneshjou R Ouyang D Ho DE Zou J How medical AI devices are evaluated: limitations and recommendations from an analysis of FDA approvals.Nat Med. 2021; 27: 582-584Crossref PubMed Scopus (182) Google Scholar These responses would not only help harness the potential advantages of using AI to predict and interpret ECGs but would also assist health-care professionals in recognising its possible limitations. This understanding will facilitate better integration with clinical practices, providing improved medical services for patients. This work was supported by the National Natural Science Foundation of China (number 62102008) and Peking University People's Hospital Scientific Research Development Funds (RDJP2022-39). We declare no competing interests. Electrocardiographic deep learning for predicting post-procedural mortality: a model development and validation studyA deep-learning algorithm interpreting preoperative ECGs can improve discrimination of postoperative mortality. The deep-learning algorithm worked equally well for risk stratification of cardiac surgeries, non-cardiac surgeries, and catheterisation laboratory procedures, and was validated in three independent health-care systems. This algorithm can provide additional information to clinicians making the decision to perform medical procedures and stratify the risk of future complications. Full-Text PDF Open Access

Safety regulation of machine learning in cardiac surgery

Machine learning methods for perioperative anesthetic management in cardiac surgery patients: a scoping review

Machine Learning Method for the Management of Acute Kidney Injury: More Than Just Treating Biomarkers Individually

Understanding Risk Factors for Postoperative Mortality in Neonates Based on Explainable Machine Learning Technology

Teaming with Artificial Intelligence to Support Global Cancer Surgical Care

Artificial intelligence in cardiac surgery: A systematic review

Machine learning for real-time prediction of complications in critical care: a retrospective study

Five Steps in Performing Machine Learning for Binary Outcomes

Expanding Electrocardiogram Abilities for Postoperative Mortality Prediction with Deep Learning

Implementing Machine Learning in Interventional Cardiology: The Benefits Are Worth the Trouble

Machine Learning in Cardio-Oncology: New Insights from an Emerging Discipline

Machine Intelligence in Cardiovascular Medicine

Effect of Machine Learning on Anaesthesiology Clinician Prediction of Postoperative Complications: The Perioperative ORACLE Randomised Clinical Trial

A Systematic Review: Do the Use of Machine Learning, Deep Learning, and Artificial Intelligence Improve Patient Outcomes in Acute Myocardial Ischemia Compared to Clinician-Only Approaches?

Overview and Clinical Applications of Artificial Intelligence and Machine Learning in Cardiac Anesthesiology

Machine learning in cardiac surgery: a narrative review

Accuracy of machine learning in predicting outcomes post-percutaneous coronary intervention: a systematic review

The Development of Machine Learning in Lung Surgery: A Narrative Review

Risk predictions of surgical wound complications based on a machine learning algorithm: A systematic review

Ethical challenges and opportunities in applying artificial intelligence to cardiovascular medicine