|Year : 2019 | Volume
| Issue : 1 | Page : 29-32
Machine learning applications in critical care
Mohammed Al Dhoayan1, Huda Alghamdi2, Yaseen M Arabi3
1 Department of Health Informatics, CPHHI, King Saud Bin Abdulaziz University for Health Sciences; Data and Business Intelligence Management Department, ISID, King Abdulaziz Medical City, Riyadh, Saudi Arabia
2 Data and Business Intelligence Management Department, ISID, King Abdulaziz Medical City, Riyadh, Saudi Arabia
3 College of Medicine, King Saud Bin Abdulaziz University for Health Sciences, King Abdullah International Medical Research Center; Department of Intensive Care, King Abdulaziz Medical City, Riyadh, Saudi Arabia
|Date of Web Publication||30-May-2019|
Mohammed Al Dhoayan
Department of Health Informatics, CPHHI, King Saud Bin Abdulaziz University for Health Sciences, Riyadh; Saudi Arabia Data and Business Intelligence Management Department, ISID, King Abdulaziz Medical City, Riyadh
Source of Support: None, Conflict of Interest: None
The use of machine learning (ML) applications in the intensive care units (ICUs) has surged over the last two decades. This is the result of the digital transformation that many health-care organizations have implemented. Data that are generated in the process of intensive care have more volume, velocity, and value than data generated in any other general hospital's department. This characteristic of ICUs makes them attractive environments for developing models that require rich dataset. ML has been used to develop clinical decision support system (CDSS) that could make informative decisions without requiring prior in-depth knowledge about the roots of the disease or common characteristics of the patients. The adoption of ML-based CDSS in ICUs is continuously increasing as ML algorithms achieve high levels of accuracy in descriptive, diagnostic, predictive, and prescriptive decisions. This article reviews some of the applications of ML in ICUs. This article will show examples of how ML was used for outcome predictions, such as predicting mortality and readmission. Examples in this article also include using ML for diagnostic and image recognition purposes. This review will discuss the use of ML for monitoring ICU patients, whether monitoring their physical safety with artificial intelligence vision detection algorithms, monitoring their continuous bedside measurements, or monitoring the administration and dosage of their medications. All these examples show that ML-based CDSS are on the path for a journey full of innovative and creative solutions that will increase the quality, efficiency, and effectiveness of critical care.
|How to cite this article:|
Al Dhoayan M, Alghamdi H, Arabi YM. Machine learning applications in critical care. Saudi Crit Care J 2019;3:29-32
| Introduction|| |
Today, health care is awash in data. With the increasing number of data sources and advancements in information technology, health-care data reached amounts of volume, velocity, variety, veracity, and value to levels higher than ever before. The intensive care unit (ICU) is a special department in health-care facilities that provides intensive care to critically ill patients. Thus, ICUs usually have a higher number of staff and resources than the other general hospital wards, which was proven to reduce rates of mortality, lower hospital length of stay, and decrease illness complications. This also means that data are collected more intensively than other departments, which gives more chance for its data to be used in descriptive, diagnostic, predictive, and prescriptive data analysis.
Nowadays, bedside equipment, such as ventilators, heart rate (HR) and electrocardiogram monitors, blood pressure (BP) and flow monitors, and infusion pumps, all store digital data and transmit them to computer databases. Electronic intensive care systems are linked with the general hospitals' electronic health record (EHR), which merges the ICU data with demographic, laboratory, radiology, and pharmacy data. Data in EHR are present in various formats including clinical documentation, imaging results, genome sequences, and continuous bedside measurements. Conventionally, data from these sources have been used for documenting clinical activities for reporting, liability, and billing reasons. Recently, a significant interest has emerged in the use of these data sources for developing machine learning (ML) algorithms and artificial intelligence (AI) models for clinical decision support systems (CDSS). The use of these systems varies from diagnostic determination and patient monitoring to the prediction of prognosis and procedural complications. Although CDSS have been used for a long time in health care, most of these tools were developed based on rules extracted from experts' knowledge or best practice guidelines. The main difference between knowledge-based CDSS and ML-based ones is that ML-based CDSS could be built on patterns of data that were never discovered before. ML-based CDSS could also revolve and relearn faster from the new data they are exposed to during their deployment. It is worth mentioning here that CDSS in general and ML-based CDSS in specific are not developed to replace health-care professionals as they have not reached the level of making decisions, they are only developed to support the process of making them. CDSS have been shown in several studies to improve the safety of patients, improve health-care quality, and reduce human medical errors.
| Machine Learning Applications for Outcome Prediction|| |
One of the early ML applications in ICUs was introduced by Tu and Guerriere where they used artificial neural network (ANN) to predict patients who would have a prolonged ICU length of stay. ANN is an ML technique that is designed to mimic the functionality of the human brain. It takes input data variables (or nodes in ANN terminology), which activate a variable number of internal layers of nodes (or hidden layers) that in turn activate the output nodes that produce the predicted value. All the nodes within the ANN are linked with edges that have different weights, according to their importance, that are adjusted and assigned to them during the training process. Tu and Guerriere used 15 input data variables from 713 patients who underwent cardiac surgery to train the ANN model and data from 696 patients for testing. The performance of their model was adequate with an area under the receiver operating characteristic curve (AUROC) of 0.70. A perfect model would score 1.00 for the AUROC, which is the measure of the tradeoff between the true positive rate and the false positive rate.
Another early ML application was developed by Doig et al. to predict mortality in ICU patients who stayed over 72 h. In their model, Doig et al. also used 15 data variables from 422 ICU patients (284 for training and 138 for testing) as input to two models, ANN, and logistic regression. These variables include but not limited to the presence of acute renal failure, packed cell volume, HR, FiO2, serum sodium, PaO2, pH, systolic and diastolic BPs, serum potassium, temperature, white blood cell count, serum creatinine, and the Glasgow Coma Score (GCS). Both models' performance was very good with an AUROC of 0.81. Doig et al. argued that both models would perform significantly better if the used dataset was larger with more variety of patients that cover more mortality reasons. They also argued that ANN will play a significant role in enhancing quality assurance, resource allocation, and the evaluation of new therapies in the near future.
The examples of early implementations of ML models, especially ANN, have sparked subsequent efforts in investigating and developing AI applications in ICUs. Recently, a decision tree (DT) model was developed by Desautels et al. to predict unplanned ICU readmission or in-hospital death within 48 h of the first ICU discharge. To build their model, Desautels et al. used patients' demographics, vital signs, and laboratory measurements as variables of interest to predict the desired outcome. After excluding missing records and underage patients, the final dataset consisted of 2018 patients who visited the ICU at some point during their stay at Cambridge University Hospitals' NHS Foundation Trust between October 2014 and August 2016. When developing DTs, one should be cautious of the potential complexity that the model could reach if not restricted with a certain depth and width level. A complicated model could introduce overfitting, which is a state when the model becomes too specific and ungeneralizable because it tried to tune its hyperparameters to perfectly fit and predict the training cases. Therefore, Desautels et al. limited their DT to split no more than eight times, and then used the Adaptive Boosting (AdaBoost) technique to generate 1000 DTs and use their aggregate votes to be the final prediction. To evaluate the performance of their model, they split the dataset into ten folds where they used nine folds for training and the remaining fold for testing in each iteration. The aggregate model evaluation of all the iterations yielded an adequate AUROC of 0.71. Desautels et al. also expressed that the small sample size was one of the challenges they faced during their analysis. Furthermore, they expressed that the generalizability of their results is limited because the predictions were made in a selected, retrospective population of patients in a single tertiary care facility. However, they argued that further analyses that follow their footsteps could help identify and devote resources to patients with unplanned readmission, which decreases variance in care, makes resource planning easier, and potentially decreases the length of stay and mortality in some settings.
| Machine Learning Applications for Clinical Diagnosis and Image Recognition|| |
In addition to using ML and AI models for outcome prediction, they are also used for the purpose of clinical diagnosis and image recognition. ML and AI applications gain higher importance in cases where traditional diagnosis measurements or tests are expensive, have a high rate of false positives or false negatives, or require special tools and setup. The power of ML algorithms is in their ability to recognize patterns in the data and their capability to cluster patients based on unobserved correlations in their characteristics. For example, disseminated intravascular coagulation (DIC) is a condition that affects a wide variety of patients, including patients admitted to ICUs. The major challenge in diagnosing DIC is the absence of biomarkers that are specific enough to be used in the diagnostic process. Although multiple composite scoring systems have been developed for this condition, optimal fibrin-related markers and their cutoff values remain to be defined. As a solution, Yoon et al. have deployed several ML models that use patients' data to diagnose DIC and compared their performance to three traditional DIC scoring systems (International Society on Thrombosis and Haemostasis [ISTH], Japanese Ministry of Health and Welfare's criteria [JMHW], and Japanese Association for Acute Medicine's criteria [JAAM]). To accomplish this task, Yoon et al. have collected 46 DIC-related variables from 656 DIC-suspected cases, half of which were ICU patients. To train the models, Yoon et al. followed the timeline of physicians' diagnostic process (reviewing clinical signs, symptoms, laboratory results, etc.) to label 228 patients as DIC and 428 as non-DIC. Using external dataset from a different hospital for validation, the ANN model outperformed all the other ML algorithms (logistic regression, linear regression, and random forest) and the three DIC scoring systems (AUROC: ANN 0.981; ISTH 0.945; JMHW 0.943; and JAAM 0.928).
The diagnostic ability of ML algorithms was also shown in the work of Sun et al., where they used deep learning to diagnose lung cancer images. In their study, they acquired 1018 samples of lung cancer images (52 × 52 pixel) each with their corresponding truth files. Four radiologists reviewed the acquired images and marked the suspicious lesions with five malignancy-level ratings. After trying different deep learning techniques, their final model reached an accuracy level of 81% using deep-belief networks.
| Machine Learning Applications for Patient Monitoring|| |
Even more, ML and AI models were not only used to diagnose patients who already have the condition of interest, several studies have implemented models that could predict and detect the conditions prior to their occurrence. Such models are very useful in ICUs to be used as monitoring tools for early detection of conditions, which permits taking precautionary actions that could prevent or reduce the harmful consequences of these conditions. Sepsis is a very common condition in ICUs and “is among the leading causes of morbidity, mortality, and cost overruns in critically ill patients.” The challenge with sepsis is that, after its confirmation, there is little time for clinicians to act before it causes serious damage to the affected patients. Nemati et al. trained a model that uses continuous bedside measurements along with other data such as laboratory results, demographics, and patient history to continuously monitor patients and raise flags or alarms when a patient is at risk of developing sepsis. In their model, they used a modified Weibull-Cox proportional hazards model, which is a survival analysis algorithm that in addition to predicting whether an sepsis instance will occur or not predicts the time of its occurrence. Their approach was to collect a dataset of 65 features; some are static that do not change with time (e.g., age, gender, and ethnicity) and some are dynamic that do change with time and arranged to have a value every hour (e.g., HR, GCS, and BP). For each record, for each patient, they calculated a Sequential Organ Failure Assessment (SOFA) score and defined a sepsis flag that is positive when an episode of suspected infection occurs with two or more point change in SOFA score. Then, to create their outcome measure, they created sliding time windows (4, 6, 8, and 12 h) that move through the records, hour by hour, to check whether an instance of sepsis has occurred during that time window or not. If a sepsis instance has occurred, all the records within that specific window will have a value of 0 for the outcome measure indicating that a sepsis event has occurred; otherwise, all records will have a value of 1 indicating that a sepsis event has not occurred. To train their model, they used data from more than 31,000 admissions to the ICUs at two Emory University hospitals. To test their model, they used data from over 52,000 ICU patients from the publicly available Medical Information Mart for Intensive Care-III database. Their model hit its highest prediction performance level when predicting sepsis 4 h prior to its occurrence (AUROC 0.85). Other time windows scored AUROC of 0.85, 0.84, and 0.83 for 6, 8, and 12, respectively.
| Machine Learning Applications for Other Purposes|| |
ML and AI are also used to solve other problems such as the ones related to medication dosing. Medication dosing in ICUs requires sophisticated decision-making based on multiple patient-related physiologic and laboratory factors. For example, ML and AI models have been developed to help in the administration of glucose homeostasis and insulin dosing using dynamic Bayesian networks, intravenous heparin dose using deep reinforcement learning, and the infusion of atracurium-induced neuromuscular block using self-learning fuzzy logic control. ML and AI are also used for image recognition tasks, such as human activity recognition for fall detection and segmentation of patient images in the neonatal ICU.
| Conclusion|| |
All these examples show that ML and AI are extremely relevant to the field of health care in general and to critical care in specific. Critical care environments are known to be rich in data because they involve intensive care and continuous patient data collection. The collection, processing, interpretation, and presentation of such data in a clinically efficient and effective format is still the primary challenge of the use of data science in ICUs. ML and AI are projected to change the structure of CDSS as they, study after study, reveal powerful insights and capabilities. ML-based CDSS will increase the personalization touch of the knowledge-based CDSS as they treat each individual uniquely, which will reduce the undesirable high number of false-positive cases that knowledge-based CDSS churn. In the near future, ML and AI applications in critical health care will witness a significant increase in their quantity and quality as many health-care organizations have reached a level of technical, in terms of the amount of stored data and computational power, and awareness maturity that enables the implementation and adaptation of such models.
Financial support and sponsorship
Conflicts of interest
There are no conflicts of interest.
| References|| |
Bailly S, Meyfroidt G, Timsit JF. What's new in ICU in 2050: Big data and machine learning. Intensive Care Med 2018;44:1524-7.
Pronovost PJ, Angus DC, Dorman T, Robinson KA, Dremsizov TT, Young TL, et al.
Physician staffing patterns and clinical outcomes in critically ill patients: A systematic review. JAMA 2002;288:2151-62.
Johnson AE, Ghassemi MM, Nemati S, Niehaus KE, Clifton DA, Clifford GD, et al.
Machine learning and decision support in critical care. Proc IEEE Inst Electr Electron Eng 2016;104:444-66.
Jao CS, Hier DB. Clinical decision support systems: An effective pathway to reduce medical errors and improve patient safety. In: Decision Support Systems. IntechOpen; 2010.
Tu JV, Guerriere MR. Use of a neural network as a predictive instrument for length of stay in the intensive care unit following cardiac surgery. Comput Biomed Res 1993;26:220-9.
Desautels T, Das R, Calvert J, Trivedi M, Summers C, Wales DJ, et al.
Prediction of early unplanned intensive care unit readmission in a UK tertiary care hospital: A cross-sectional machine learning approach. BMJ Open 2017;7:e017199.
Yoon JG, Heo J, Kim M, Park YJ, Choi MH, Song J, et al.
Machine learning-based diagnosis for disseminated intravascular coagulation (DIC): Development, external validation, and comparison to scoring systems. PLoS One 2018;13:e0195861.
Sun W, Zheng B, Qian W. Computer aided lung cancer diagnosis with deep learning algorithms. In: Medical Imaging 2016: Computer-Aided Diagnosis. Vol. 9785. 2016. p. 97850Z.
Nemati S, Holder A, Razmi F, Stanley MD, Clifford GD, Buchman TG, et al.
An interpretable machine learning model for accurate prediction of sepsis in the ICU. Crit Care Med 2018;46:547-53.
Nachimuthu SK, Wong A, Haug PJ. Modeling glucose homeostasis and insulin dosing in an intensive care unit using dynamic Bayesian networks. AMIA Annu Symp Proc 2010;2010:532-6.
Lin R, Stanley MD, Ghassemi MM, Nemati S. A deep deterministic policy gradient approach to medication dosing and surveillance in the ICU. Conf Proc IEEE Eng Med Biol Soc 2018;2018:4927-31.
Mason DG, Ross JJ, Edwards ND, Linkens DA, Reilly CS. Self-learning fuzzy control of atracurium-induced neuromuscular block during surgery. Med Biol Eng Comput 1997;35:498-503.
Diraco G, Leone A, Siciliano P. An active vision system for fall detection and posture recognition in elderly healthcare. In: Proceedings of the Conference on Design, Automation and Test in Europe. 2010. p. 1536-41.
Dossso YS, Bekele A, Nizami S, Aubertin C, Greenwood K, Harrold J, et al
. Segmentation of patient images in the neonatal intensive care unit. In: 2018 IEEE Life Sciences Conference (LSC). 2018. p. 45-8.
Vellido A, Ribas V, Morales C, Ruiz Sanmartín A, Ruiz Rodríguez JC. Machine learning in critical care: State-of-the-art and a sepsis case study. Biomed Eng Online 2018;17:135.