Using Supervised Machine Learning to Classify Real Alerts and Artifact in Online Multisignal Vital Sign Monitoring Data. Crit Care Med. 2016;44:e456-e463. [PMID: 26992068 DOI: 10.1097/ccm.0000000000001660]

New sensors for the early detection of clinical deterioration on general wards and beyond - a clinician's perspective

The early detection of clinical deterioration could be the next significant step in enhancing patient safety in general hospital wards. Most patients do not deteriorate suddenly; instead, their vital signs are often abnormal or trending towards an abnormal range hours before severe adverse events requiring rescue intervention and/or ICU transfer. To date, at least 10 large clinical studies have demonstrated a significant reduction in severe adverse events when heart rate, blood pressure, oxygen saturation and/or respiratory rate are continuously monitored on medical and surgical wards. Continuous, silent, and automatic monitoring of vital signs also presents the opportunity to eliminate unnecessary spot-checks for stable patients. This could lead to a reduction in nurse workload, while significantly improving patient comfort, sleep quality, and overall satisfaction. Wireless and wearable sensors are particularly valuable, as they make continuous monitoring feasible even for ambulatory patients, raising questions about the future relevance of "stay-in-bed" solutions like capnography, bed sensors, and video-monitoring systems. While the number of wearable sensors and mobile monitoring solutions is rapidly growing, independent validation studies on their sensitivity and specificity in detecting abnormal vital signs in actual patients, rather than healthy volunteers, remain limited. Additionally, further research is needed to evaluate the cost-effectiveness of using wireless wearables for vital sign monitoring both within hospital wards and at home.

A Comparison of Five Algorithmic Methods and Machine Learning Pattern Recognition for Artifact Detection in Electronic Records of Five Different Vital Signs: A Retrospective Analysis

BACKGROUND

Research on electronic health record physiologic data is common, invariably including artifacts. Traditionally, these artifacts have been handled using simple filter techniques. The authors hypothesized that different artifact detection algorithms, including machine learning, may be necessary to provide optimal performance for various vital signs and clinical contexts.

METHODS

In a retrospective single-center study, intraoperative operating room and intensive care unit (ICU) electronic health record datasets including heart rate, oxygen saturation, blood pressure, temperature, and capnometry were included. All records were screened for artifacts by at least two human experts. Classical artifact detection methods (cutoff, multiples of SD [z-value], interquartile range, and local outlier factor) and a supervised learning model implementing long short-term memory neural networks were tested for each vital sign against the human expert reference dataset. For each artifact detection algorithm, sensitivity and specificity were calculated.

RESULTS

A total of 106 (53 operating room and 53 ICU) patients were randomly selected, resulting in 392,808 data points. Human experts annotated 5,167 (1.3%) data points as artifacts. The artifact detection algorithms demonstrated large variations in performance. The specificity was above 90% for all detection methods and all vital signs. The neural network showed significantly higher sensitivities than the classic methods for heart rate (ICU, 33.6%; 95% CI, 33.1 to 44.6), systolic invasive blood pressure (in both the operating room [62.2%; 95% CI, 57.5 to 71.9] and the ICU [60.7%; 95% CI, 57.3 to 71.8]), and temperature in the operating room (76.1%; 95% CI, 63.6 to 89.7). The CI for specificity overlapped for all methods. Generally, sensitivity was low, with only the z-value for oxygen saturation in the operating room reaching 88.9%. All other sensitivities were less than 80%.

CONCLUSIONS

No single artifact detection method consistently performed well across different vital signs and clinical settings. Neural networks may be a promising artifact detection method for specific vital signs.

EDITOR’S PERSPECTIVE

Artificial Intelligence and Machine Learning Applications in Critically Ill Brain Injured Patients

The utilization of Artificial Intelligence (AI) and Machine Learning (ML) is paving the way for significant strides in patient diagnosis, treatment, and prognostication in neurocritical care. These technologies offer the potential to unravel complex patterns within vast datasets ranging from vast clinical data and EEG (electroencephalogram) readings to advanced cerebral imaging facilitating a more nuanced understanding of patient conditions. Despite their promise, the implementation of AI and ML faces substantial hurdles. Historical biases within training data, the challenge of interpreting multifaceted data streams, and the "black box" nature of ML algorithms present barriers to widespread clinical adoption. Moreover, ethical considerations around data privacy and the need for transparent, explainable models remain paramount to ensure trust and efficacy in clinical decision-making.This article reflects on the emergence of AI and ML as integral tools in neurocritical care, discussing their roles from the perspective of both their scientific promise and the associated challenges. We underscore the importance of extensive validation in diverse clinical settings to ensure the generalizability of ML models, particularly considering their potential to inform critical medical decisions such as withdrawal of life-sustaining therapies. Advancement in computational capabilities is essential for implementing ML in clinical settings, allowing for real-time analysis and decision support at the point of care. As AI and ML are poised to become commonplace in clinical practice, it is incumbent upon health care professionals to understand and oversee these technologies, ensuring they adhere to the highest safety standards and contribute to the realization of personalized medicine. This engagement will be pivotal in integrating AI and ML into patient care, optimizing outcomes in neurocritical care through informed and data-driven decision-making.

Artificial Intelligence in Perioperative Planning and Management of Liver Resection

Artificial intelligence (AI) is a speciality within computer science that deals with creating systems that can replicate the intelligence of a human mind and has problem-solving abilities. AI includes a diverse array of techniques and approaches such as machine learning, neural networks, natural language processing, robotics, and expert systems. An electronic literature search was conducted using the databases of "PubMed" and "Google Scholar". The period for the search was from 2000 to June 2023. The search terms included "artificial intelligence", "machine learning", "liver cancers", "liver tumors", "hepatectomy", "perioperative" and their synonyms in various combinations. The search also included all MeSH terms. The extracted articles were further reviewed in a step-wise manner for identification of relevant studies. A total of 148 articles were identified after the initial literature search. Initial review included screening of article titles for relevance and identifying duplicates. Finally, 65 articles were reviewed for this review article. The future of AI in liver cancer planning and management holds immense promise. AI-driven advancements will increasingly enable precise tumour detection, location, and characterisation through enhanced image analysis. ML algorithms will predict patient-specific treatment responses and complications, allowing for tailored therapies. Surgical robots and AI-guided procedures will enhance the precision of liver resections, reducing risks and improving outcomes. AI will also streamline patient monitoring, better hemodynamic management, enabling early detection of recurrence or complications. Moreover, AI will facilitate data-driven research, accelerating the development of novel treatments and therapies. Ultimately, AI's integration will revolutionise liver cancer care, offering personalised, efficient and effective solutions, improving patients' quality of life and survival rates.

Continuous monitoring of physiological data using the patient vital status fusion score in septic critical care patients

Accurate and standardized methods for assessing the vital status of patients are crucial for patient care and scientific research. This study introduces the Patient Vital Status (PVS), which quantifies and contextualizes a patient's physical status based on continuous variables such as vital signs and deviations from age-dependent normative values. The vital signs, heart rate, oxygen saturation, respiratory rate, mean arterial blood pressure, and temperature were selected as input to the PVS pipeline. The method was applied to 70 pediatric patients in the intensive care unit (ICU), and its efficacy was evaluated by matching high values with septic events at different time points in patient care. Septic events included systemic inflammatory response syndrome (SIRS) and suspected or proven sepsis. The comparison of maximum PVS values between the presence and absence of a septic event showed significant differences (SIRS/No SIRS: p < 0.0001, η2 = 0.54; Suspected Sepsis/No Suspected Sepsis: p = 0.00047, η2 = 0.43; Proven Sepsis/No Proven Sepsis: p = 0.0055, η2 = 0.34). A further comparison between the most severe PVS in septic patients with the PVS at ICU discharge showed even higher effect sizes (SIRS: p < 0.0001, η2 = 0.8; Suspected Sepsis: p < 0.0001, η2 = 0.8; Proven Sepsis: p = 0.002, η2 = 0.84). The PVS is emerging as a data-driven tool with the potential to assess a patient's vital status in the ICU objectively. Despite real-world data challenges and potential annotation biases, it shows promise for monitoring disease progression and treatment responses. Its adaptability to different disease markers and reliance on age-dependent reference values further broaden its application possibilities. Real-time implementation of PVS in personalized patient monitoring may be a promising way to improve critical care. However, PVS requires further research and external validation to realize its true potential.

Interdisciplinary collaboration in critical care alarm research: A bibliometric analysis

BACKGROUND

Alarm fatigue in nurses is a major patient safety concern in the intensive care unit. This is caused by exposure to high rates of false and non-actionable alarms. Despite decades of research, the problem persists, leading to stress, burnout, and patient harm resulting from true missed events. While engineering approaches to reduce false alarms have spurred hope, they appear to lack collaboration between nurses and engineers to produce real-world solutions. The aim of this bibliometric analysis was to examine the relevant literature to quantify the level of authorial collaboration between nurses, physicians, and engineers.

METHODS

We conducted a bibliometric analysis of articles on alarm fatigue and false alarm reduction strategies in critical care published between 2010 and 2022. Data were extracted at the article and author level. The percentages of author disciplines per publication were calculated by study design, journal subject area, and other article-level factors.

RESULTS

A total of 155 articles with 583 unique authors were identified. While 31.73 % (n = 185) of the unique authors had a nursing background, publications using an engineering study design (n = 46), e.g., model development, had a very low involvement of nursing authors (mean proportion at 1.09 %). Observational studies (n = 58) and interventional studies (n = 33) had a higher mean involvement of 52.27 % and 47.75 %, respectively. Articles published in nursing journals (n = 32) had the highest mean proportion of nursing authors (80.32 %), while those published in engineering journals (n = 46) had the lowest (9.00 %), with 6 (13.04 %) articles having one or more nurses as co-authors.

CONCLUSION

Minimal involvement of nursing expertise in alarm research utilizing engineering methodologies may be one reason for the lack of successful, real-world solutions to ameliorate alarm fatigue. Fostering a collaborative, interdisciplinary research culture can promote a common publication culture across fields and may yield sustainable implementation of technological solutions in healthcare.

Weakly Supervised Classification of Vital Sign Alerts as Real or Artifact

A significant proportion of clinical physiologic monitoring alarms are false. This often leads to alarm fatigue in clinical personnel, inevitably compromising patient safety. To combat this issue, researchers have attempted to build Machine Learning (ML) models capable of accurately adjudicating Vital Sign (VS) alerts raised at the bedside of hemodynamically monitored patients as real or artifact. Previous studies have utilized supervised ML techniques that require substantial amounts of hand-labeled data. However, manually harvesting such data can be costly, time-consuming, and mundane, and is a key factor limiting the widespread adoption of ML in healthcare (HC). Instead, we explore the use of multiple, individually imperfect heuristics to automatically assign probabilistic labels to unlabeled training data using weak supervision. Our weakly supervised models perform competitively with traditional supervised techniques and require less involvement from domain experts, demonstrating their use as efficient and practical alternatives to supervised learning in HC applications of ML.

Computational approaches to alleviate alarm fatigue in intensive care medicine: A systematic literature review

Patient monitoring technology has been used to guide therapy and alert staff when a vital sign leaves a predefined range in the intensive care unit (ICU) for decades. However, large amounts of technically false or clinically irrelevant alarms provoke alarm fatigue in staff leading to desensitisation towards critical alarms. With this systematic review, we are following the Preferred Reporting Items for Systematic Reviews (PRISMA) checklist in order to summarise scientific efforts that aimed to develop IT systems to reduce alarm fatigue in ICUs. 69 peer-reviewed publications were included. The majority of publications targeted the avoidance of technically false alarms, while the remainder focused on prediction of patient deterioration or alarm presentation. The investigated alarm types were mostly associated with heart rate or arrhythmia, followed by arterial blood pressure, oxygen saturation, and respiratory rate. Most publications focused on the development of software solutions, some on wearables, smartphones, or headmounted displays for delivering alarms to staff. The most commonly used statistical models were tree-based. In conclusion, we found strong evidence that alarm fatigue can be alleviated by IT-based solutions. However, future efforts should focus more on the avoidance of clinically non-actionable alarms which could be accelerated by improving the data availability.

Systematic Review Registration:https://www.crd.york.ac.uk/prospero/display_record.php?ID=CRD42021233461, identifier: CRD42021233461.

Wireless wearables for postoperative surveillance on surgical wards: a survey of 1158 anaesthesiologists in Western Europe and the USA

Background

Several continuous monitoring solutions, including wireless wearable sensors, are available or being developed to improve patient surveillance on surgical wards. We designed a survey to understand the current perception and expectations of anaesthesiologists who, as perioperative physicians, are increasingly involved in postoperative care.

Methods

The survey was shared in 40 university hospitals from Western Europe and the USA.

Results

From 5744 anaesthesiologists who received the survey link, there were 1158 valid questionnaires available for analysis. Current postoperative surveillance was mainly based on intermittent spot-checks of vital signs every 4-6 h in the USA (72%) and every 8-12 h in Europe (53%). A majority of respondents (91%) considered that continuous monitoring of vital signs should be available on surgical wards and that wireless sensors are preferable to tethered systems (86%). Most respondents indicated that oxygen saturation (93%), heart rate (80%), and blood pressure (71%) should be continuously monitored with wrist devices (71%) or skin adhesive patches (54%). They believed it may help detect clinical deterioration earlier (90%), decrease rescue interventions (59%), and decrease hospital mortality (54%). Opinions diverged regarding the impact on nurse workload (increase 46%, decrease 39%), and most respondents considered that the biggest implementation challenges are economic (79%) and connectivity issues (64%).

Conclusion

Continuous monitoring of vital signs with wireless sensors is wanted by most anaesthesiologists from university hospitals in Western Europe and in the USA. They believe it may improve patient safety and outcome, but may also be challenging to implement because of cost and connectivity issues.

Using Machine Learning to Support Transfer of Best Practices in Healthcare

The adoption of best practices has been shown to increase performance in healthcare institutions and is consistently demanded by both patients, payers, and external overseers. Nevertheless, transferring practices between healthcare organizations is a challenging and underexplored task. In this paper, we take a step towards enabling the transfer of best practices by identifying the likely beneficial opportunities for such transfer. Specifically, we analyze the output of machine learning models trained at different organizations with the aims of (i) detecting the opportunity for the transfer of best practices, and (ii) providing a stop-gap solution while the actual transfer process takes place. We show the benefits ofthis methodology on a dataset ofmedical inpatient claims, demonstrating our abilityto identify practice gaps and to support the transfer processes that address these gaps.

Construction and Comparison of Predictive Models for Length of Stay after Total Knee Arthroplasty: Regression Model and Machine Learning Analysis Based on 1,826 Cases in a Single Singapore Center

The purpose of this study was to develop a predictive model for length of stay (LOS) after total knee arthroplasty (TKA). Between 2013 and 2014, 1,826 patients who underwent TKA from a single Singapore center were enrolled in the study after qualification. Demographics of patients with normal and prolonged LOS were analyzed. The risk variables that could affect LOS were identified by univariate analysis. Predictive models for LOS after TKA by logistic regression or machine learning were constructed and compared. The univariate analysis showed that age, American Society of Anesthesiologist level, diabetes, ischemic heart disease, congestive heart failure, general anesthesia, and operation duration were risk factors that could affect LOS (p < 0.05). Comparing with logistic regression models, the machine learning model with all variables was the best model to predict LOS after TKA, of whose area of operator characteristic curve was 0.738. Machine learning algorithms improved the predictive performance of LOS prediction models for TKA patients.

Nurses' use of auditory alarms and alerts in high dependency units: A field study

A fieldwork study conducted in six units of a major metropolitan Australian hospital revealed that nurses' attitudes towards alarms are influenced by each unit's physical layout and caseload. Additionally, nurses relied heavily on both non-actionable and actionable alarms to maintain their awareness of the status of their patients' wellbeing, and used auditory alarms beyond the scope of their intended design. Results suggest that before reducing or removing auditory alarms from the clinical environment to improve patient safety, it is important to understand how nurses in different clinical contexts use current alarm systems to extract meaningful information. Such an understanding could guide appropriate alarm reduction strategies and guide alternative design solutions to support nurses' situation awareness during monitoring.

[Machine learning in anesthesiology]

The application of artificial intelligence (AI) is currently changing very different areas of life. Artificial intelligence involves the emulation of human behavior with the aid of methods from mathematics and informatics. Machine learning (ML) represents a subdivision of AI. Algorithms for ML have the potential to optimize patient care, in that they can be utilized in a supportive way in personalized medicine, decision making and risk prediction. Although the majority of the applications in medicine are still limited to data analysis and research, it is certain that ML will become increasingly more important in scientific and clinical aspects in this supportive function. Therefore, it is necessary for clinicians to have at least a basic understanding of the functional principles, strengths and weaknesses of ML.

Patient Deterioration in the Adult Progressive Care Unit: A Scoping Review

BACKGROUND

The American Heart Association set a goal in 2010 to double the percentage of hospitalized adult patients who survive cardiac arrest by 2020. Because of acuity and interventions, progressive care patients are a population of interest to address this goal. The state of the literature involving patient deterioration, which can lead to cardiac arrest, in the progressive care setting has yet to be explored.

OBJECTIVE

A scoping review was done to investigate the literature involving patient deterioration in adult progressive care units in order to map knowledge, identify themes, and discover areas for research potential.

METHODS

The scoping review began with an extensive literature search and a multistep review. The characteristics of the final group of studies were charted and grouped according to common themes.

RESULTS

There were 13 studies in the final group. All studies were conducted in the United States and most by interprofessional teams. Three themes were evident in the review, training methods, surveillance, and monitoring systems.

DISCUSSION

Patient deterioration in the progressive care unit may benefit from team-based training methods involving checklists or protocols. Nurses can use surveillance, including physical assessment and technology, to recognize early warning signs. Lastly, the use of systems that identify patterns in vital signs can be useful to reduce patient harm. Further research in this area care is warranted and could potentially improve patient outcomes and nursing practice.

A machine learning approach to predict healthcare-associated infections at intensive care unit admission: findings from the SPIN-UTI project

BACKGROUND

Identifying patients at higher risk of healthcare-associated infections (HAIs) in intensive care units (ICUs) represents a major challenge for public health. Machine learning could improve patient risk stratification and lead to targeted infection prevention and control interventions.

AIM

To evaluate the performance of the Simplified Acute Physiology Score (SAPS) II for HAI risk prediction in ICUs, using both traditional statistical and machine learning approaches.

METHODS

Data for 7827 patients from the 'Italian Nosocomial Infections Surveillance in Intensive Care Units' project were used in this study. The Support Vector Machines (SVM) algorithm was applied to classify patients according to sex, patient origin, non-surgical treatment for acute coronary disease, surgical intervention, SAPS II at admission, presence of invasive devices, trauma, impaired immunity, and antibiotic therapy in 48 h preceding ICU admission.

FINDINGS

The performance of SAPS II for predicting HAI risk provides a receiver operating characteristic curve with an area under the curve of 0.612 (P<0.001) and accuracy of 56%. Considering SAPS II along with other characteristics at ICU admission, the SVM classifier was found to have accuracy of 88% and an AUC of 0.90 (P<0.001) for the test set. The predictive ability was lower when considering the same SVM model but with the SAPS II variable removed (accuracy 78%, AUC 0.66).

CONCLUSIONS

This study suggested that the SVM model is a useful tool for early prediction of patients at higher risk of HAIs at ICU admission.

Future Medical Artificial Intelligence Application Requirements and Expectations of Physicians in German University Hospitals: Web-Based Survey

BACKGROUND

The increasing development of artificial intelligence (AI) systems in medicine driven by researchers and entrepreneurs goes along with enormous expectations for medical care advancement. AI might change the clinical practice of physicians from almost all medical disciplines and in most areas of health care. While expectations for AI in medicine are high, practical implementations of AI for clinical practice are still scarce in Germany. Moreover, physicians' requirements and expectations of AI in medicine and their opinion on the usage of anonymized patient data for clinical and biomedical research have not been investigated widely in German university hospitals.

OBJECTIVE

This study aimed to evaluate physicians' requirements and expectations of AI in medicine and their opinion on the secondary usage of patient data for (bio)medical research (eg, for the development of machine learning algorithms) in university hospitals in Germany.

METHODS

A web-based survey was conducted addressing physicians of all medical disciplines in 8 German university hospitals. Answers were given using Likert scales and general demographic responses. Physicians were asked to participate locally via email in the respective hospitals.

RESULTS

The online survey was completed by 303 physicians (female: 121/303, 39.9%; male: 173/303, 57.1%; no response: 9/303, 3.0%) from a wide range of medical disciplines and work experience levels. Most respondents either had a positive (130/303, 42.9%) or a very positive attitude (82/303, 27.1%) towards AI in medicine. There was a significant association between the personal rating of AI in medicine and the self-reported technical affinity level (H4=48.3, P<.001). A vast majority of physicians expected the future of medicine to be a mix of human and artificial intelligence (273/303, 90.1%) but also requested a scientific evaluation before the routine implementation of AI-based systems (276/303, 91.1%). Physicians were most optimistic that AI applications would identify drug interactions (280/303, 92.4%) to improve patient care substantially but were quite reserved regarding AI-supported diagnosis of psychiatric diseases (62/303, 20.5%). Of the respondents, 82.5% (250/303) agreed that there should be open access to anonymized patient databases for medical and biomedical research.

CONCLUSIONS

Physicians in stationary patient care in German university hospitals show a generally positive attitude towards using most AI applications in medicine. Along with this optimism comes several expectations and hopes that AI will assist physicians in clinical decision making. Especially in fields of medicine where huge amounts of data are processed (eg, imaging procedures in radiology and pathology) or data are collected continuously (eg, cardiology and intensive care medicine), physicians' expectations of AI to substantially improve future patient care are high. In the study, the greatest potential was seen in the application of AI for the identification of drug interactions, assumedly due to the rising complexity of drug administration to polymorbid, polypharmacy patients. However, for the practical usage of AI in health care, regulatory and organizational challenges still have to be mastered.

Near-Infrared Spectroscopy to Assess Cerebral Autoregulation and Optimal Mean Arterial Pressure in Patients With Hypoxic-Ischemic Brain Injury: A Prospective Multicenter Feasibility Study

Supplemental Digital Content is available in the text.

We provide preliminary multicenter data to suggest that recruitment and collection of physiologic data necessary to quantify cerebral autoregulation and individualized blood pressure targets are feasible in postcardiac arrest patients. We evaluated the feasibility of a multicenter protocol to enroll patients across centers, as well as collect continuous recording (≥ 80% of monitoring time) of regional cerebral oxygenation and mean arterial pressure, which is required to quantify cerebral autoregulation, using the cerebral oximetry index, and individualized optimal mean arterial pressure thresholds. Additionally, we conducted an exploratory analysis to assess if an increased percentage of monitoring time where mean arterial pressure was greater than or equal to 5 mm Hg below optimal mean arterial pressure, percentage of monitoring time with dysfunctional cerebral autoregulation (i.e., cerebral oximetry index ≥ 0.3), and time to return of spontaneous circulation were associated with an unfavorable neurologic outcome (i.e., 6-mo Cerebral Performance Category score ≥ 3).

Effects of monitor alarm management training on nurses' alarm fatigue: A randomised controlled trial

BACKGROUND

Chaotic monitor alarm management generates a large number of alarms, which result in alarm fatigue. Intensive care unit (ICU) nurses are caretakers of critically ill patients, the effect of alarm management affect patient safety directly.

OBJECTIVES

To evaluate the effect of monitor alarm management training based on the theory of planned behaviour for reducing alarm fatigue in intensive care unit nurses.

DESIGN

A randomised, single-blind trial. This article follows the requirements of CONSORT statement.

PARTICIPANTS

The study was conducted from February 2019-May 2019 in a tertiary A-level hospital. 93 ICU clinical nurses were included, and they were randomly assigned into two groups.

INTERVENTION

Nurses in the experimental group (n = 47) received a 12-week alarm management training course based on the theory of planned behaviour. Nurses in the control group (n = 46) received regular training. All nurses' alarm fatigue scores were measured with a questionnaire before and after the study period. Total number of alarms, nonactionable alarms and true crisis alarms were recorded continuously throughout the study period.

RESULTS

For baseline comparisons, no significant differences were found. By the analysis of independent samples one-way ANCOVAs, the nurses' adjusted alarm fatigue scores at the post-test in the experimental group were significantly lower than those in the control group (p < .001). After the study period, adjusted total number of alarms and nonactionable alarms recorded in the experimental group were both significantly lower than those recorded in the control group (p < .001). After the study period, no significant difference between the two groups was noted in the adjusted number of true crisis alarms (p > .05). The interventions did not cause adverse events in either group of patients and did not cause adverse events in patients.

CONCLUSION

Intensive care unit nurses' alarm fatigue was effectively decreased by the monitor alarm management training based on the theory of planned behaviour.

RELEVANCE TO CLINICAL PRACTICE

(1) Monitor alarm training based on the theory of planned behaviour is effective in reducing nonactionable alarms and lowering alarm fatigue in ICU nurses. (2) The intervention considering the social psychological aspects of behaviour is effective in rebuilding the nurses' awareness and behaviour of alarm management. (3) Nurses are the direct users of monitoring technology. Hospital administrators should attach importance to the role of nurses in the medical monitoring system. We suggest that nursing managers implement training programmes in more ICUs in the future to improve alarm management ability and lower alarm fatigue in ICU nurses.

Artifacts annotations in anesthesia blood pressure data by man and machine

Physiologic data from anesthesia monitors are automatically captured. Yet erroneous data are stored in the process as well. While this is not interfering with clinical care, research can be affected. Researchers should find ways to remove artifacts. The aim of the present study was to compare different artifact annotation strategies, and to assess if a machine learning algorithm is able to accept or reject individual data points. Non-cardiac procedures requiring invasive blood pressure monitoring were eligible. Two trained research assistants observed procedures live for artifacts. The same procedures were also retrospectively annotated for artifacts by a different person. We compared the different ways of artifact identifications and modelled artifacts with three different learning algorithms (lasso restrictive logistic regression, neural network and support vector machine). In 88 surgical procedures including 5711 blood pressure data points, the live observed incidence of artifacts was 2.1% and the retrospective incidence was 2.2%. Comparing retrospective with live annotation revealed a sensitivity of 0.32 and specificity of 0.98. The performance of the learning algorithms which we applied ranged from poor (kappa 0.053) to moderate (kappa 0.651). Manual identification of artifacts yielded different incidences in different situations, which were not comparable. Artifact detection in physiologic data collected during anesthesia could be automated, but the performance of the learning algorithms in the present study remained moderate. Future research should focus on optimization and finding ways to apply them with minimal manual work. The present study underlines the importance of an explicit definition for artifacts in database research.

Role of Artificial Intelligence in Patient Safety Outcomes: Systematic Literature Review

BACKGROUND

Artificial intelligence (AI) provides opportunities to identify the health risks of patients and thus influence patient safety outcomes.

OBJECTIVE

The purpose of this systematic literature review was to identify and analyze quantitative studies utilizing or integrating AI to address and report clinical-level patient safety outcomes.

METHODS

We restricted our search to the PubMed, PubMed Central, and Web of Science databases to retrieve research articles published in English between January 2009 and August 2019. We focused on quantitative studies that reported positive, negative, or intermediate changes in patient safety outcomes using AI apps, specifically those based on machine-learning algorithms and natural language processing. Quantitative studies reporting only AI performance but not its influence on patient safety outcomes were excluded from further review.

RESULTS

We identified 53 eligible studies, which were summarized concerning their patient safety subcategories, the most frequently used AI, and reported performance metrics. Recognized safety subcategories were clinical alarms (n=9; mainly based on decision tree models), clinical reports (n=21; based on support vector machine models), and drug safety (n=23; mainly based on decision tree models). Analysis of these 53 studies also identified two essential findings: (1) the lack of a standardized benchmark and (2) heterogeneity in AI reporting.

CONCLUSIONS

This systematic review indicates that AI-enabled decision support systems, when implemented correctly, can aid in enhancing patient safety by improving error detection, patient stratification, and drug management. Future work is still needed for robust validation of these systems in prospective and real-world clinical environments to understand how well AI can predict safety outcomes in health care settings.

Artifact Processing Methods Influence on Intraoperative Hypotension Quantification and Outcome Effect Estimates

BACKGROUND

Physiologic data that is automatically collected during anesthesia is widely used for medical record keeping and clinical research. These data contain artifacts, which are not relevant in clinical care, but may influence research results. The aim of this study was to explore the effect of different methods of filtering and processing artifacts in anesthesiology data on study findings in order to demonstrate the importance of proper artifact filtering.

METHODS

The authors performed a systematic literature search to identify artifact filtering methods. Subsequently, these methods were applied to the data of anesthesia procedures with invasive blood pressure monitoring. Different hypotension measures were calculated (i.e., presence, duration, maximum deviation below threshold, and area under threshold) across different definitions (i.e., thresholds for mean arterial pressure of 50, 60, 65, 70 mmHg). These were then used to estimate the association with postoperative myocardial injury.

RESULTS

After screening 3,585 papers, the authors included 38 papers that reported artifact filtering methods. The authors applied eight of these methods to the data of 2,988 anesthesia procedures. The occurrence of hypotension (defined with a threshold of 50 mmHg) varied from 24% with a median filter of seven measurements to 55% without an artifact filtering method, and between 76 and 90% with a threshold of 65 mmHg. Standardized odds ratios for presence of hypotension ranged from 1.16 (95% CI, 1.07 to 1.26) to 1.24 (1.14 to 1.34) when hypotension was defined with a threshold of 50 mmHg. Similar variations in standardized odds ratios were found when applying methods to other hypotension measures and definitions.

CONCLUSIONS

The method of artifact filtering can have substantial effects on estimates of hypotension prevalence. The effect on the association between intraoperative hypotension and postoperative myocardial injury was relatively small. Nevertheless, the authors recommend that researchers carefully consider artifacts handling and report the methodology used.

Artificial intelligence and computer simulation models in critical illness

Widespread implementation of electronic health records has led to the increased use of artificial intelligence (AI) and computer modeling in clinical medicine. The early recognition and treatment of critical illness are central to good outcomes but are made difficult by, among other things, the complexity of the environment and the often non-specific nature of the clinical presentation. Increasingly, AI applications are being proposed as decision supports for busy or distracted clinicians, to address this challenge. Data driven "associative" AI models are built from retrospective data registries with missing data and imprecise timing. Associative AI models lack transparency, often ignore causal mechanisms, and, while potentially useful in improved prognostication, have thus far had limited clinical applicability. To be clinically useful, AI tools need to provide bedside clinicians with actionable knowledge. Explicitly addressing causal mechanisms not only increases validity and replicability of the model, but also adds transparency and helps gain trust from the bedside clinicians for real world use of AI models in teaching and patient care.

Development, Implementation, and Evaluation of a Personalized Machine Learning Algorithm for Clinical Decision Support: Case Study With Shingles Vaccination

BACKGROUND

Although clinical decision support (CDS) alerts are effective reminders of best practices, their effectiveness is blunted by clinicians who fail to respond to an overabundance of inappropriate alerts. An electronic health record (EHR)-integrated machine learning (ML) algorithm is a potentially powerful tool to increase the signal-to-noise ratio of CDS alerts and positively impact the clinician's interaction with these alerts in general.

OBJECTIVE

This study aimed to describe the development and implementation of an ML-based signal-to-noise optimization system (SmartCDS) to increase the signal of alerts by decreasing the volume of low-value herpes zoster (shingles) vaccination alerts.

METHODS

We built and deployed SmartCDS, which builds personalized user activity profiles to suppress shingles vaccination alerts unlikely to yield a clinician's interaction. We extracted all records of shingles alerts from January 2017 to March 2019 from our EHR system, including 327,737 encounters, 780 providers, and 144,438 patients.

RESULTS

During the 6 weeks of pilot deployment, the SmartCDS system suppressed an average of 43.67% (15,425/35,315) potential shingles alerts (appointments) and maintained stable counts of weekly shingles vaccination orders (326.3 with system active vs 331.3 in the control group; P=.38) and weekly user-alert interactions (1118.3 with system active vs 1166.3 in the control group; P=.20).

CONCLUSIONS

All key statistics remained stable while the system was turned on. Although the results are promising, the characteristics of the system can be subject to future data shifts, which require automated logging and monitoring. We demonstrated that an automated, ML-based method and data architecture to suppress alerts are feasible without detriment to overall order rates. This work is the first alert suppression ML-based model deployed in practice and serves as foundational work in encounter-level customization of alert display to maximize effectiveness.

Opportunities for machine learning to improve surgical ward safety

BACKGROUND

Delayed recognition of decompensation and failure-to-rescue on surgical wards are major sources of preventable harm. This review assimilates and critically evaluates available evidence and identifies opportunities to improve surgical ward safety.

DATA SOURCES

Fifty-eight articles from Cochrane Library, EMBASE, and PubMed databases were included.

CONCLUSIONS

Only 15-20% of patients suffering ward arrest survive. In most cases, subtle signs of instability often occur prior to critical illness and arrest, and underlying pathology is reversible. Coarse risk assessments lead to under-triage of high-risk patients to wards, where surveillance for complications depends on time-consuming manual review of health records, infrequent patient assessments, prediction models that lack accuracy and autonomy, and biased, error-prone decision-making. Streaming electronic heath record data, wearable continuous monitors, and recent advances in deep learning and reinforcement learning can promote efficient and accurate risk assessments, earlier recognition of instability, and better decisions regarding diagnosis and treatment of reversible underlying pathology.

Predictive Monitoring of Critical Cardiorespiratory Alarms in Neonates Under Intensive Care

We aimed at reducing alarm fatigue in neonatal intensive care units by developing a model using machine learning for the early prediction of critical cardiorespiratory alarms. During this study in over 34,000 patient monitoring hours in 55 infants 278,000 advisory (yellow) and 70,000 critical (red) alarms occurred. Vital signs including the heart rate, breathing rate, and oxygen saturation were obtained at a sampling frequency of 1 Hz while heart rate variability was calculated by processing the ECG - both were used for feature development and for predicting alarms. Yellow alarms that were followed by at least one red alarm within a short post-alarm window constituted the case-cohort while the remaining yellow alarms constituted the control cohort. For analysis, the case and control cohorts, stratified by proportion, were split into training (80%) and test sets (20%). Classifiers based on decision trees were used to predict, at the moment the yellow alarm occurred, whether a red alarm(s) would shortly follow. The best performing classifier used data from the 2-min window before the occurrence of the yellow alarm and could predict 26% of the red alarms in advance (18.4s, median), at the expense of 7% additional red alarms. These results indicate that based on predictive monitoring of critical alarms, nurses can be provided a longer window of opportunity for preemptive clinical action. Further, such as algorithm can be safely implemented as alarms that are not algorithmically predicted can still be generated upon the usual breach of the threshold, as in current clinical practice.

Increasing Cardiovascular Data Sampling Frequency and Referencing It to Baseline Improve Hemorrhage Detection

Supplemental Digital Content is available in the text.

We hypothesize that knowledge of a stable personalized baseline state and increased data sampling frequency would markedly improve the ability to detect progressive hypovolemia during hemorrhage earlier and with a lower false positive rate than when using less granular data.

Unexpected sawtooth artifact in beat-to-beat pulse transit time measured from patient monitor data

Object

It is increasingly popular to collect as much data as possible in the hospital setting from clinical monitors for research purposes. However, in this setup the data calibration issue is often not discussed and, rather, implicitly assumed, while the clinical monitors might not be designed for the data analysis purpose. We hypothesize that this calibration issue for a secondary analysis may become an important source of artifacts in patient monitor data. We test an off-the-shelf integrated photoplethysmography (PPG) and electrocardiogram (ECG) monitoring device for its ability to yield a reliable pulse transit time (PTT) signal.

Approach

This is a retrospective clinical study using two databases: one containing 35 subjects who underwent laparoscopic cholecystectomy, another containing 22 subjects who underwent spontaneous breathing test in the intensive care unit. All data sets include recordings of PPG and ECG using a commonly deployed patient monitor. We calculated the PTT signal offline.

Main results

We report a novel constant oscillatory pattern in the PTT signal and identify this pattern as a sawtooth artifact. We apply an approach based on the de-shape method to visualize, quantify and validate this sawtooth artifact.

Significance

The PPG and ECG signals not designed for the PTT evaluation may contain unwanted artifacts. The PTT signal should be calibrated before analysis to avoid erroneous interpretation of its physiological meaning.

Signal quality measure for pulsatile physiological signals using morphological features: Applications in reliability measure for pulse oximetry

Pulse oximetry is a noninvasive and low-cost physiological monitor that measures blood oxygen levels. While the noninvasive nature of pulse oximetry is advantageous, the estimates of oxygen saturation generated by these devices are prone to motion artifacts and ambient noise, reducing the reliability of such estimations. Clinicians combat this by assessing the quality of oxygen saturation estimation by visual inspection of the photoplethysmograph (PPG), which represents changes in pulsatile blood volume and is also generated by the pulse oximeter. In this paper, we propose six morphological features that can be used to determine the quality of the PPG signal and generate a signal quality index. Unlike many similar studies, this approach uses machine learning and does not require a separate signal, such as ECG, for reference. Multiple algorithms were tested against 46 30-min PPG segments of patients with cardiovascular and respiratory conditions, including atrial fibrillation, hypoxia, acute heart failure, pneumonia, ARDS, and pulmonary embolism. These signals were independently annotated for signal quality by two clinicians, with the union of their annotations used as the ground-truth. Similar to any physiological signal recorded in a clinical setting, the utilized dataset is also unbalanced in favor of good quality segments. The experiments showed that a cost-sensitive Support Vector Machine (SVM) outperformed other tested methods and was robust to the unbalanced nature of the data. Though the proposed algorithm was tested on PPG signals, the methodology remains agnostic to the dataset used, and may be applied to any type of pulsatile physiological signal.

Machine-learning Algorithm to Predict Hypotension Based on High-fidelity Arterial Pressure Waveform Analysis

WHAT WE ALREADY KNOW ABOUT THIS TOPIC

WHAT THIS ARTICLE TELLS US THAT IS NEW: BACKGROUND:: With appropriate algorithms, computers can learn to detect patterns and associations in large data sets. The authors' goal was to apply machine learning to arterial pressure waveforms and create an algorithm to predict hypotension. The algorithm detects early alteration in waveforms that can herald the weakening of cardiovascular compensatory mechanisms affecting preload, afterload, and contractility.

METHODS

The algorithm was developed with two different data sources: (1) a retrospective cohort, used for training, consisting of 1,334 patients' records with 545,959 min of arterial waveform recording and 25,461 episodes of hypotension; and (2) a prospective, local hospital cohort used for external validation, consisting of 204 patients' records with 33,236 min of arterial waveform recording and 1,923 episodes of hypotension. The algorithm relates a large set of features calculated from the high-fidelity arterial pressure waveform to the prediction of an upcoming hypotensive event (mean arterial pressure < 65 mmHg). Receiver-operating characteristic curve analysis evaluated the algorithm's success in predicting hypotension, defined as mean arterial pressure less than 65 mmHg.

RESULTS

Using 3,022 individual features per cardiac cycle, the algorithm predicted arterial hypotension with a sensitivity and specificity of 88% (85 to 90%) and 87% (85 to 90%) 15 min before a hypotensive event (area under the curve, 0.95 [0.94 to 0.95]); 89% (87 to 91%) and 90% (87 to 92%) 10 min before (area under the curve, 0.95 [0.95 to 0.96]); 92% (90 to 94%) and 92% (90 to 94%) 5 min before (area under the curve, 0.97 [0.97 to 0.98]).

CONCLUSIONS

The results demonstrate that a machine-learning algorithm can be trained, with large data sets of high-fidelity arterial waveforms, to predict hypotension in surgical patients' records.

Automated continuous noninvasive ward monitoring: future directions and challenges

Automated continuous noninvasive ward monitoring may enable subtle changes in vital signs to be recognized. There is already some evidence that automated ward monitoring can improve patient outcome. Before automated continuous noninvasive ward monitoring can be implemented in clinical routine, several challenges and problems need to be considered and resolved; these include the meticulous validation of the monitoring systems with regard to their measurement performance, minimization of artifacts and false alarms, integration and combined analysis of massive amounts of data including various vital signs, and technical problems regarding the connectivity of the systems.

The physiological determinants of near-infrared spectroscopy-derived regional cerebral oxygenation in critically ill adults

Background

To maintain adequate oxygen delivery to tissue, resuscitation of critically ill patients is guided by assessing surrogate markers of perfusion. As there is no direct indicator of cerebral perfusion used in routine critical care, identifying an accurate strategy to monitor brain perfusion is paramount. Near-infrared spectroscopy (NIRS) is a non-invasive technique to quantify regional cerebral oxygenation (rSO₂) that has been used for decades during cardiac surgery which has led to targeted algorithms to optimize rSO₂ being developed. However, these targeted algorithms do not exist during critical care, as the physiological determinants of rSO₂ during critical illness remain poorly understood.

Materials and methods

This prospective observational study was an exploratory analysis of a nested cohort of patients within the CONFOCAL study (NCT02344043) who received high-fidelity vital sign monitoring. Adult patients (≥ 18 years) admitted < 24 h to a medical/surgical intensive care unit were eligible if they had shock and/or required mechanical ventilation. Patients underwent rSO₂ monitoring with the FORESIGHT oximeter for 24 h, vital signs were concurrently recorded, and clinically ordered arterial blood gas samples and hemoglobin concentration were also documented. Simultaneous multiple linear regression was performed using all available predictors, followed by model selection using the corrected Akaike information criterion (AICc).

Results

Our simultaneous multivariate model included age, heart rate, arterial oxygen saturation, mean arterial pressure, pH, partial pressure of oxygen, partial pressure of carbon dioxide (PaCO₂), and hemoglobin concentration. This model accounted for a significant proportion of variance in rSO₂ (R² = 0.58, p < 0.01) and was significantly associated with PaCO₂ (p < 0.05) and hemoglobin concentration (p < 0.01). Our selected regression model using AICc accounted for a significant proportion of variance in rSO₂ (R² = 0.54, p < 0.01) and was significantly related to age (p < 0.05), PaCO₂ (p < 0.01), hemoglobin (p < 0.01), and heart rate (p < 0.05).

Conclusions

Known and established physiological determinants of oxygen delivery accounted for a significant proportion of the rSO₂ signal, which provides evidence that NIRS is a viable modality to assess cerebral oxygenation in critically ill adults. Further elucidation of the determinants of rSO₂ has the potential to develop a NIRS-guided resuscitation algorithm during critical illness.

Trial registration

This trial is registered on clinicaltrials.gov (Identifier: NCT02344043), retrospectively registered January 8, 2015.

Electronic supplementary material

The online version of this article (10.1186/s40635-019-0247-0) contains supplementary material, which is available to authorized users.

Predicting adverse hemodynamic events in critically ill patients

PURPOSE OF REVIEW

The art of predicting future hemodynamic instability in the critically ill has rapidly become a science with the advent of advanced analytical processed based on computer-driven machine learning techniques. How these methods have progressed beyond severity scoring systems to interface with decision-support is summarized.

RECENT FINDINGS

Data mining of large multidimensional clinical time-series databases using a variety of machine learning tools has led to our ability to identify alert artifact and filter it from bedside alarms, display real-time risk stratification at the bedside to aid in clinical decision-making and predict the subsequent development of cardiorespiratory insufficiency hours before these events occur. This fast evolving filed is primarily limited by linkage of high-quality granular to physiologic rationale across heterogeneous clinical care domains.

SUMMARY

Using advanced analytic tools to glean knowledge from clinical data streams is rapidly becoming a reality whose clinical impact potential is great.

Intensive Care Unit Telemedicine in the Era of Big Data, Artificial Intelligence, and Computer Clinical Decision Support Systems

This article examines the history of the telemedicine intensive care unit (tele-ICU), the current state of clinical decision support systems (CDSS) in the tele-ICU, applications of machine learning (ML) algorithms to critical care, and opportunities to integrate ML with tele-ICU CDSS. The enormous quantities of data generated by tele-ICU systems is a major driver in the development of the large, comprehensive, heterogeneous, and granular data sets necessary to develop generalizable ML CDSS algorithms, and deidentification of these data sets expands opportunities for ML CDSS research.

Predicting tachycardia as a surrogate for instability in the intensive care unit

Tachycardia is a strong though non-specific marker of cardiovascular stress that proceeds hemodynamic instability. We designed a predictive model of tachycardia using multi-granular intensive care unit (ICU) data by creating a risk score and dynamic trajectory. A subset of clinical and numerical signals were extracted from the Multiparameter Intelligent Monitoring in Intensive Care II database. A tachycardia episode was defined as heart rate ≥ 130/min lasting for ≥ 5 min, with ≥ 10% density. Regularized logistic regression (LR) and random forest (RF) classifiers were trained to create a risk score for upcoming tachycardia. Three different risk score models were compared for tachycardia and control (non-tachycardia) groups. Risk trajectory was generated from time windows moving away at 1 min increments from the tachycardia episode. Trajectories were computed over 3 hours leading up to the episode for three different models. From 2809 subjects, 787 tachycardia episodes and 707 control periods were identified. Patients with tachycardia had increased vasopressor support, longer ICU stay, and increased ICU mortality than controls. In model evaluation, RF was slightly superior to LR, which accuracy ranged from 0.847 to 0.782, with area under the curve from 0.921 to 0.842. Risk trajectory analysis showed average risks for tachycardia group evolved to 0.78 prior to the tachycardia episodes, while control group risks remained < 0.3. Among the three models, the internal control model demonstrated evolving trajectory approximately 75 min before tachycardia episode. Clinically relevant tachycardia episodes can be predicted from vital sign time series using machine learning algorithms.

Hemodynamic Instability and Cardiovascular Events After Traumatic Brain Injury Predict Outcome After Artifact Removal With Deep Belief Network Analysis

BACKGROUND

Hemodynamic instability and cardiovascular events heavily affect the prognosis of traumatic brain injury. Physiological signals are monitored to detect these events. However, the signals are often riddled with faulty readings, which jeopardize the reliability of the clinical parameters obtained from the signals. A machine-learning model for the elimination of artifactual events shows promising results for improving signal quality. However, the actual impact of the improvements on the performance of the clinical parameters after the elimination of the artifacts is not well studied.

MATERIALS AND METHODS

The arterial blood pressure of 99 subjects with traumatic brain injury was continuously measured for 5 consecutive days, beginning on the day of admission. The machine-learning deep belief network was constructed to automatically identify and remove false incidences of hypotension, hypertension, bradycardia, tachycardia, and alterations in cerebral perfusion pressure (CPP).

RESULTS

The prevalences of hypotension and tachycardia were significantly reduced by 47.5% and 13.1%, respectively, after suppressing false incidents (P=0.01). Hypotension was particularly effective at predicting outcome favorability and mortality after artifact elimination (P=0.015 and 0.027, respectively). In addition, increased CPP was also statistically significant in predicting outcomes (P=0.02).

CONCLUSIONS

The prevalence of false incidents due to signal artifacts can be significantly reduced using machine-learning. Some clinical events, such as hypotension and alterations in CPP, gain particularly high predictive capacity for patient outcomes after artifacts are eliminated from physiological signals.

Applying machine learning to continuously monitored physiological data

The use of machine learning (ML) in healthcare has enormous potential for improving disease detection, clinical decision support, and workflow efficiencies. In this commentary, we review published and potential applications for the use of ML for monitoring within the hospital environment. We present use cases as well as several questions regarding the application of ML to the analysis of the vast amount of complex data that clinicians must interpret in the realm of continuous physiological monitoring. ML, especially employed in bidirectional conjunction with electronic health record data, has the potential to extract much more useful information out of this currently under-analyzed data source from a population level. As a data driven entity, ML is dependent on copious, high quality input data so that error can be introduced by low quality data sources. At present, while ML is being studied in hybrid formulations along with static expert systems for monitoring applications, it is not yet actively incorporated in the formal artificial learning sense of an algorithm constantly learning and updating its rules without external intervention. Finally, innovations in monitoring, including those supported by ML, will pose regulatory and medico-legal challenges, as well as questions regarding precisely how to incorporate these features into clinical care and medical education. Rigorous evaluation of ML techniques compared to traditional methods or other AI methods will be required to validate the algorithms developed with consideration of database limitations and potential learning errors. Demonstration of value on processes and outcomes will be necessary to support the use of ML as a feature in monitoring system development: Future research is needed to evaluate all AI based programs before clinical implementation in non-research settings.

A call to alarms: Current state and future directions in the battle against alarm fatigue

Research demonstrates that the majority of alarms derived from continuous bedside monitoring devices are non-actionable. This avalanche of unreliable alerts causes clinicians to experience sensory overload when attempting to sort real from false alarms, causing desensitization and alarm fatigue, which in turn leads to adverse events when true instability is neither recognized nor attended to despite the alarm. The scope of the problem of alarm fatigue is broad, and its contributing mechanisms are numerous. Current and future approaches to defining and reacting to actionable and non-actionable alarms are being developed and investigated, but challenges in impacting alarm modalities, sensitivity and specificity, and clinical activity in order to reduce alarm fatigue and adverse events remain. A multi-faceted approach involving clinicians, computer scientists, industry, and regulatory agencies is needed to battle alarm fatigue.

Towards an automated multimodal clinical decision support system at the post anesthesia care unit

BACKGROUND

The aim of this study was to develop a predictive algorithm detecting early signs of deterioration (ESODs) in the post anesthesia care unit (PACU), thus being able to intervene earlier in the future to avoid serious adverse events. The algorithm must utilize continuously collected cardiopulmonary vital signs and may serve as an alternative to current practice, in which an alarm is activated by single parameters.

METHODS

The study was a single center, prospective cohort study including 178 patients admitted to the PACU after major surgical procedures. Peripheral blood oxygenation, arterial blood pressure, perfusion index, heart rate and respiratory rate were monitored continuously. Potential ESODs were automatically detected and scored by two independent experts with regards to the severity of the observation. Based on features extracted from the obtained measurements, a random forest classifier was trained, classifying each event being either an ESOD or not an ESOD. The algorithm was evaluated and compared to the automated single modality alarm system at the PACU.

RESULTS

The algorithm detected ESODs with an accuracy of 92.2% (99% CI: 89.6%-94.8%), sensitivity of 90.6% (99% CI: 85.7%-95.5%), specificity of 93.0% (99% CI: 89.9%-96.2%) and area under the receiver operating characteristic curve of 96.9% (99% CI: 95.3%-98.5%). The number of false alarms decreased by 85% (99% CI: 77%-93%) and the number of missed ESODs decreased by 73% (99% CI: 61%-85%) as compared to the currently used alarm system in the hospital. The algorithm was able to detect an ESOD in average 26.4 (99% CI: 1.1-51.7) minutes before the current single parameter system used in the PACU.

CONCLUSION

In conclusion, the proposed biomedical classification algorithm, when compared to the currently used single parameter alarm system of the hospital, showed significantly increased performance in both detecting ESODs fast and classifying these correctly. The clinical effect of the predictive system must be evaluated in future trials.

Predicting Mortality in the Surgical Intensive Care Unit Using Artificial Intelligence and Natural Language Processing of Physician Documentation

The purpose of this study was to use natural language processing of physician documentation to predict mortality in patients admitted to the surgical intensive care unit (SICU). The Multiparameter Intelligent Monitoring in Intensive Care III database was used to obtain SICU stays with six different severity of illness scores. Natural language processing was performed on the physician notes. Classifiers for predicting mortality were created. One classifier used only the physician notes, one used only the severity of illness scores, and one used the physician notes with severity of injury scores. There were 3838 SICU stays identified during the study period and 5.4 per cent ended with mortality. The classifier trained with physician notes with severity of injury scores performed with the highest area under the curve (0.88 ± 0.05) and accuracy (94.6 ± 1.1%). The most important variable was the Oxford Acute Severity of Illness Score (16.0%). The most important terms were “dilated” (4.3%) and “hemorrhage” (3.7%). This study demonstrates the novel use of artificial intelligence to process physician documentation to predict mortality in the SICU. The classifiers were able to detect the subtle nuances in physician vernacular that predict mortality. These nuances provided improved performance in predicting mortality over physiologic parameters alone.

Secondary brain injury: Predicting and preventing insults

Mortality or severe disability affects the majority of patients after severe traumatic brain injury (TBI). Adherence to the brain trauma foundation guidelines has overall improved outcomes; however, traditional as well as novel interventions towards intracranial hypertension and secondary brain injury have come under scrutiny after series of negative randomized controlled trials. In fact, it would not be unfair to say there has been no single major breakthrough in the management of severe TBI in the last two decades. One plausible hypothesis for the aforementioned failures is that by the time treatment is initiated for neuroprotection, or physiologic optimization, irreversible brain injury has already set in. We, and others, have recently developed predictive models based on machine learning from continuous time series of intracranial pressure and partial brain tissue oxygenation. These models provide accurate predictions of physiologic crises events in a timely fashion, offering the opportunity for an earlier application of targeted interventions. In this article, we review the rationale for prediction, discuss available predictive models with examples, and offer suggestions for their future prospective testing in conjunction with preventive clinical algorithms. This article is part of the Special Issue entitled "Novel Treatments for Traumatic Brain Injury".

Using artificial intelligence to predict prolonged mechanical ventilation and tracheostomy placement

BACKGROUND

Early identification of critically ill patients who will require prolonged mechanical ventilation (PMV) has proven to be difficult. The purpose of this study was to use machine learning to identify patients at risk for PMV and tracheostomy placement.

MATERIALS AND METHODS

The Multiparameter Intelligent Monitoring in Intensive Care III database was queried for all intensive care unit (ICU) stays with mechanical ventilation. PMV was defined as ventilation >7 d. Classifiers with a gradient-boosted decision trees algorithm were created for the outcomes of PMV and tracheostomy placement. The variables used were six different severity-of-illness scores calculated on the first day of ICU admission including their components and 30 comorbidities. Mean receiver operating characteristic curves were calculated for the outcomes, and variable importance was quantified.

RESULTS

There were 20,262 ICU stays identified. PMV was required in 13.6%, and tracheostomy was performed in 6.6% of patients. The classifier for predicting PMV was able to achieve a mean area under the curve (AUC) of 0.820 ± 0.016, and tracheostomy was predicted with an AUC of 0.830 ± 0.011. There were 60.7% patients admitted to a surgical ICU, and the classifiers for these patients predicted PMV with an AUC of 0.852 ± 0.017 and tracheostomy with an AUC of 0.869 ± 0.015. The variable with the highest importance for predicting PMV was the logistic organ dysfunction score pulmonary component (13%), and the most important comorbidity in predicting tracheostomy was cardiac arrhythmia (12%).

CONCLUSIONS

This study demonstrates the use of artificial intelligence through machine-learning classifiers for the early identification of patients at risk for PMV and tracheostomy. Application of these identification techniques could lead to improved outcomes by allowing for early intervention.

Predicting central line-associated bloodstream infections and mortality using supervised machine learning

PURPOSE

The purpose of this study was to compare machine learning techniques for predicting central line-associated bloodstream infection (CLABSI).

MATERIALS AND METHODS

The Multiparameter Intelligent Monitoring in Intensive Care III database was queried for all ICU admissions. The variables included six different severities of illness scores calculated on the first day of ICU admission with their components and comorbidities. The outcomes of interest were in-hospital mortality, central line placement, and CLABSI. Predictive models were created for these outcomes using classifiers with different algorithms: logistic regression, gradient boosted trees, and deep learning.

RESULTS

There were 57,786 total hospital admissions and the mortality rate was 10.1%. There were 38.4% patients with a central line and the rate of CLABSI was 1.5%. The classifiers using deep learning performed with the highest AUC for mortality, 0.885±0.010 (p<0.01) and central line placement, 0.816±0.006 (p<0.01). The classifier using logistic regression for predicting CLABSI performed with an AUC of 0.722±0.048 (p<0.01).

CONCLUSIONS

This study demonstrates models for identifying patients who will develop CLABSI. Early identification of these patients has implications for quality, cost, and outcome improvements.

Dynamic and Personalized Risk Forecast in Step-Down Units. Implications for Monitoring Paradigms

RATIONALE

Cardiorespiratory insufficiency (CRI) is a term applied to the manifestations of loss of normal cardiorespiratory reserve and portends a bad outcome. CRI occurs commonly in hospitalized patients, but its risk escalation patterns are unexplored.

OBJECTIVES

To describe the dynamic and personal character of CRI risk evolution observed through continuous vital sign monitoring of individual step-down unit patients.

METHODS

Using a machine learning model, we estimated risk trends for CRI (defined as exceedance of vital sign stability thresholds) for each of 1,971 admissions (1,880 unique patients) to a 24-bed adult surgical trauma step-down unit at an urban teaching hospital in Pittsburgh, Pennsylvania using continuously recorded vital signs from standard bedside monitors. We compared and contrasted risk trends during initial 4-hour periods after step-down unit admission, and again during the 4 hours immediately before the CRI event, between cases (ever had a CRI) and control subjects (never had a CRI). We further explored heterogeneity of risk escalation patterns during the 4 hours before CRI among cases, comparing personalized to nonpersonalized risk.

MEASUREMENTS AND MAIN RESULTS

Estimated risk was significantly higher for cases (918) than control subjects (1,053; P ≤ 0.001) during the initial 4-hour stable periods. Among cases, the aggregated nonpersonalized risk trend increased 2 hours before the CRI, whereas the personalized risk trend became significantly different from control subjects 90 minutes ahead. We further discovered several unique phenotypes of risk escalation patterns among cases for nonpersonalized (14.6% persistently high risk, 18.6% early onset, 66.8% late onset) and personalized risk (7.7% persistently high risk, 8.9% early onset, 83.4% late onset).

CONCLUSIONS

Insights from this proof-of-concept analysis may guide design of dynamic and personalized monitoring systems that predict CRI, taking into account the triage and real-time monitoring utility of vital signs. These monitoring systems may prove useful in the dynamic allocation of technological and clinical personnel resources in acute care hospitals.