Search for: All records

Abstract Accurate prediction of postoperative complications can inform shared decisions regarding prognosis, preoperative risk-reduction, and postoperative resource use. We hypothesized that multi-task deep learning models would outperform conventional machine learning models in predicting postoperative complications, and that integrating high-resolution intraoperative physiological time series would result in more granular and personalized health representations that would improve prognostication compared to preoperative predictions. In a longitudinal cohort study of 56,242 patients undergoing 67,481 inpatient surgical procedures at a university medical center, we compared deep learning models with random forests and XGBoost for predicting nine common postoperative complications using preoperative, intraoperative, and perioperative patient data. Our study indicated several significant results across experimental settings that suggest the utility of deep learning for capturing more precise representations of patient health for augmented surgical decision support. Multi-task learning improved efficiency by reducing computational resources without compromising predictive performance. Integrated gradients interpretability mechanisms identified potentially modifiable risk factors for each complication. Monte Carlo dropout methods provided a quantitative measure of prediction uncertainty that has the potential to enhance clinical trust. Multi-task learning, interpretability mechanisms, and uncertainty metrics demonstrated potential to facilitate effective clinical implementation. 

Transformer model architectures have revolutionized the natural language processing (NLP) domain and continue to produce state-of-the-art results in text-based applications. Prior to the emergence of transformers, traditional NLP models such as recurrent and convolutional neural networks demonstrated promising utility for patient-level predictions and health forecasting from longitudinal datasets. However, to our knowledge only few studies have explored transformers for predicting clinical outcomes from electronic health record (EHR) data, and in our estimation, none have adequately derived a health-specific tokenization scheme to fully capture the heterogeneity of EHR systems. In this study, we propose a dynamic method for tokenizing both discrete and continuous patient data, and present a transformer-based classifier utilizing a joint embedding space for integrating disparate temporal patient measurements. We demonstrate the feasibility of our clinical AI framework through multi-task ICU patient acuity estimation, where we simultaneously predict six mortality and readmission outcomes. Our longitudinal EHR tokenization and transformer modeling approaches resulted in more accurate predictions compared with baseline machine learning models, which suggest opportunities for future multimodal data integrations and algorithmic support tools using clinical transformer networks. 

During the early stages of hospital admission, clinicians use limited information to make decisions as patient acuity evolves. We hypothesized that clustering analysis of vital signs measured within six hours of hospital admission would reveal distinct patient phenotypes with unique pathophysiological signatures and clinical outcomes. We created a longitudinal electronic health record dataset for 75,762 adult patient admissions to a tertiary care center in 2014–2016 lasting six hours or longer. Physiotypes were derived via unsupervised machine learning in a training cohort of 41,502 patients applying consensus k -means clustering to six vital signs measured within six hours of admission. Reproducibility and correlation with clinical biomarkers and outcomes were assessed in validation cohort of 17,415 patients and testing cohort of 16,845 patients. Training, validation, and testing cohorts had similar age (54–55 years) and sex (55% female), distributions. There were four distinct clusters. Physiotype A had physiologic signals consistent with early vasoplegia, hypothermia, and low-grade inflammation and favorable short-and long-term clinical outcomes despite early, severe illness. Physiotype B exhibited early tachycardia, tachypnea, and hypoxemia followed by the highest incidence of prolonged respiratory insufficiency, sepsis, acute kidney injury, and short- and long-term mortality. Physiotype C had minimal early physiological derangement and favorable clinical outcomes. Physiotype D had the greatest prevalence of chronic cardiovascular and kidney disease, presented with severely elevated blood pressure, and had good short-term outcomes but suffered increased 3-year mortality. Comparing sequential organ failure assessment (SOFA) scores across physiotypes demonstrated that clustering did not simply recapitulate previously established acuity assessments. In a heterogeneous cohort of hospitalized patients, unsupervised machine learning techniques applied to routine, early vital sign data identified physiotypes with unique disease categories and distinct clinical outcomes. This approach has the potential to augment understanding of pathophysiology by distilling thousands of disease states into a few physiological signatures. 

Mistrust is a major barrier to implementing deep learning in healthcare settings. Entrustment could be earned by conveying model certainty, or the probability that a given model output is accurate, but the use of uncertainty estimation for deep learning entrustment is largely unexplored, and there is no consensus regarding optimal methods for quantifying uncertainty. Our purpose is to critically evaluate methods for quantifying uncertainty in deep learning for healthcare applications and propose a conceptual framework for specifying certainty of deep learning predictions. We searched Embase, MEDLINE, and PubMed databases for articles relevant to study objectives, complying with PRISMA guidelines, rated study quality using validated tools, and extracted data according to modified CHARMS criteria. Among 30 included studies, 24 described medical imaging applications. All imaging model architectures used convolutional neural networks or a variation thereof. The predominant method for quantifying uncertainty was Monte Carlo dropout, producing predictions from multiple networks for which different neurons have dropped out and measuring variance across the distribution of resulting predictions. Conformal prediction offered similar strong performance in estimating uncertainty, along with ease of interpretation and application not only to deep learning but also to other machine learning approaches. Among the six articles describing non-imaging applications, model architectures and uncertainty estimation methods were heterogeneous, but predictive performance was generally strong, and uncertainty estimation was effective in comparing modeling methods. Overall, the use of model learning curves to quantify epistemic uncertainty (attributable to model parameters) was sparse. Heterogeneity in reporting methods precluded the performance of a meta-analysis. Uncertainty estimation methods have the potential to identify rare but important misclassifications made by deep learning models and compare modeling methods, which could build patient and clinician trust in deep learning applications in healthcare. Efficient maturation of this field will require standardized guidelines for reporting performance and uncertainty metrics. 

Established guidelines describe minimum requirements for reporting algorithms in healthcare; it is equally important to objectify the characteristics of ideal algorithms that confer maximum potential benefits to patients, clinicians, and investigators. We propose a framework for ideal algorithms, including 6 desiderata: explainable (convey the relative importance of features in determining outputs), dynamic (capture temporal changes in physiologic signals and clinical events), precise (use high-resolution, multimodal data and aptly complex architecture), autonomous (learn with minimal supervision and execute without human input), fair (evaluate and mitigate implicit bias and social inequity), and reproducible (validated externally and prospectively and shared with academic communities). We present an ideal algorithms checklist and apply it to highly cited algorithms. Strategies and tools such as the predictive, descriptive, relevant (PDR) framework, the Standard Protocol Items: Recommendations for Interventional Trials-Artificial Intelligence (SPIRIT-AI) extension, sparse regression methods, and minimizing concept drift can help healthcare algorithms achieve these objectives, toward ideal algorithms in healthcare. 

Accurate prediction and monitoring of patient health in the intensive care unit can inform shared decisions regarding appropriateness of care delivery, risk-reduction strategies, and intensive care resource use. Traditionally, algorithmic solutions for patient outcome prediction rely solely on data available from electronic health records (EHR). In this pilot study, we explore the benefits of augmenting existing EHR data with novel measurements from wrist-worn activity sensors as part of a clinical environment known as the Intelligent ICU. We implemented temporal deep learning models based on two distinct sources of patient data: (1) routinely measured vital signs from electronic health records, and (2) activity data collected from wearable sensors. As a proxy for illness severity, our models predicted whether patients leaving the intensive care unit would be successfully or unsuccessfully discharged from the hospital. We overcome the challenge of small sample size in our prospective cohort by applying deep transfer learning using EHR data from a much larger cohort of traditional ICU patients. Our experiments quantify added utility of non-traditional measurements for predicting patient health, especially when applying a transfer learning procedure to small novel Intelligent ICU cohorts of critically ill patients. 

Critical care patients experience varying levels of pain during their stay in the intensive care unit, often requiring administration of analgesics and sedation. Such medications generally exacerbate the already sedentary physical activity profiles of critical care patients, contributing to delayed recovery. Thus, it is important not only to minimize pain levels, but also to optimize analgesic strategies in order to maximize mobility and activity of ICU patients. Currently, we lack an understanding of the relation between pain and physical activity on a granular level. In this study, we examined the relationship between nurse assessed pain scores and physical activity as measured using a wearable accelerometer device. We found that average, standard deviation, and maximum physical activity counts are significantly higher before high pain reports compared to before low pain reports during both daytime and nighttime, while percentage of time spent immobile was not significantly different between the two pain report groups. Clusters detected among patients using extracted physical activity features were significant in adjusted logistic regression analysis for prediction of pain report group.