BACKGROUND: Venovenous extracorporeal membrane oxygenation (VV-ECMO) is a therapy for patients with refractory respiratory failure. The decision to decannulate someone from extracorporeal membrane oxygenation (ECMO) often involves weaning trials and clinical intuition. To date, there are limited prognostication metrics to guide clinical decision-making to determine which patients will be successfully weaned and decannulated.
OBJECTIVE: This study aims to assist clinicians with the decision to decannulate a patient from ECMO, using Continuous Evaluation of VV-ECMO Outcomes (CEVVO), a deep learning-based model for predicting success of decannulation in patients supported on VV-ECMO. The running metric may be applied daily to categorize patients into high-risk and low-risk groups. Using these data, providers may consider initiating a weaning trial based on their expertise and CEVVO.
METHODS: Data were collected from 118 patients supported with VV-ECMO at the Columbia University Irving Medical Center. Using a long short-term memory-based network, CEVVO is the first model capable of integrating discrete clinical information with continuous data collected from an ECMO device. A total of 12 sets of 5-fold cross validations were conducted to assess the performance, which was measured using the area under the receiver operating characteristic curve (AUROC) and average precision (AP). To translate the predicted values into a clinically useful metric, the model results were calibrated and stratified into risk groups, ranging from 0 (high risk) to 3 (low risk). To further investigate the performance edge of CEVVO, 2 synthetic data sets were generated using Gaussian process regression. The first data set preserved the long-term dependency of the patient data set, whereas the second did not.
RESULTS: CEVVO demonstrated consistently superior classification performance compared with contemporary models (P<.001 and P=.04 compared with the next highest AUROC and AP). Although the model\'s patient-by-patient predictive power may be too low to be integrated into a clinical setting (AUROC 95% CI 0.6822-0.7055; AP 95% CI 0.8515-0.8682), the patient risk classification system displayed greater potential. When measured at 72 hours, the high-risk group had a successful decannulation rate of 58% (7/12), whereas the low-risk group had a successful decannulation rate of 92% (11/12; P=.04). When measured at 96 hours, the high- and low-risk groups had a successful decannulation rate of 54% (6/11) and 100% (9/9), respectively (P=.01). We hypothesized that the improved performance of CEVVO was owing to its ability to efficiently capture transient temporal patterns. Indeed, CEVVO exhibited improved performance on synthetic data with inherent temporal dependencies (P<.001) compared with logistic regression and a dense neural network.
CONCLUSIONS: The ability to interpret and integrate large data sets is paramount for creating accurate models capable of assisting clinicians in risk stratifying patients supported on VV-ECMO. Our framework may guide future incorporation of CEVVO into more comprehensive intensive care monitoring systems.

BACKGROUND: In the hospital setting, frailty is a significant risk factor, but difficult to measure in clinical practice. We propose a reweighting of an existing diagnoses-based frailty score using routine data from a tertiary care teaching hospital in southern Germany.
METHODS: The dataset includes patient characteristics such as sex, age, primary and secondary diagnoses and in-hospital mortality. Based on this information, we recalculate the existing Hospital Frailty Risk Score. The cohort includes patients aged ≥ 75 and was divided into a development cohort (admission year 2011 to 2013, N = 30,525) and a validation cohort (2014, N = 11,202). A limited external validation is also conducted in a second validation cohort containing inpatient cases aged ≥ 75 in 2022 throughout Germany (N = 491,251). In the development cohort, LASSO regression analysis was used to select the most relevant variables and to generate a reweighted Frailty Score for the German setting. Discrimination is assessed using the area under the receiver operating characteristic curve (AUC). Visualization of calibration curves and decision curve analysis were carried out. Applicability of the reweighted Frailty Score in a non-elderly population was assessed using logistic regression models.
RESULTS: Reweighting of the Frailty Score included only 53 out of the 109 frailty-related diagnoses and resulted in substantially better discrimination than the initial weighting of the score (AUC = 0.89 vs. AUC = 0.80, p < 0.001 in the validation cohort). Calibration curves show a good agreement between score-based predictions and actual observed mortality. Additional external validation using inpatient cases aged ≥ 75 in 2022 throughout Germany (N = 491,251) confirms the results regarding discrimination and calibration and underlines the geographic and temporal validity of the reweighted Frailty Score. Decision curve analysis indicates that the clinical usefulness of the reweighted score as a general decision support tool is superior to the initial version of the score. Assessment of the applicability of the reweighted Frailty Score in a non-elderly population (N = 198,819) shows that discrimination is superior to the initial version of the score (AUC = 0.92 vs. AUC = 0.87, p < 0.001). In addition, we observe a fairly age-stable influence of the reweighted Frailty Score on in-hospital mortality, which does not differ substantially for women and men.
CONCLUSIONS: Our data indicate that the reweighted Frailty Score is superior to the original Frailty Score for identification of older, frail patients at risk for in-hospital mortality. Hence, we recommend using the reweighted Frailty Score in the German in-hospital setting.

Early-stage Alzheimer\'s disease (AD) and frontotemporal dementia (FTD) share similar symptoms, complicating their diagnosis and the development of specific treatment strategies. Our study evaluated multiple feature extraction techniques for identifying AD and FTD biomarkers from electroencephalographic (EEG) signals. We developed an optimised machine learning architecture that integrates sliding windowing, feature extraction, and supervised learning to distinguish between AD and FTD patients, as well as from healthy controls (HCs). Our model, with a 90% overlap for sliding windowing, SVD entropy for feature extraction, and K-Nearest Neighbors (KNN) for supervised learning, achieved a mean F1-score and accuracy of 93% and 91%, 92.5% and 93%, and 91.5% and 91% for discriminating AD and HC, FTD and HC, and AD and FTD, respectively. The feature importance array, an explainable AI feature, highlighted the brain lobes that contributed to identifying and distinguishing AD and FTD biomarkers. This research introduces a novel framework for detecting and discriminating AD and FTD using EEG signals, addressing the need for accurate early-stage diagnostics. Furthermore, a comparative evaluation of sliding windowing, multiple feature extraction, and machine learning methods on AD/FTD detection and discrimination is documented.

Falls by the elderly pose considerable health hazards, leading not only to physical harm but a number of other related problems. A timely alert about a deteriorating gait, as an indication of an impending fall, can assist in fall prevention. In this investigation, a comprehensive comparative analysis was conducted between a commercially available mobile phone system and two wristband systems: one commercially available and another representing a novel approach. Each system was equipped with a singular three-axis accelerometer. The walk suggestive of a potential fall was induced by special glasses worn by the participants. The same standard machine-learning techniques were employed for the classification with all three systems based on a single three-axis accelerometer, yielding a best average accuracy of 86%, a specificity of 88%, and a sensitivity of 86% via the support vector machine (SVM) method using a wristband. A smartphone, on the other hand, achieved a best average accuracy of 73% also with an SVM using only a three-axis accelerometer sensor. The significance analysis of the mean accuracy, sensitivity, and specificity between the innovative wristband and the smartphone yielded a p-value of 0.000. Furthermore, the study applied unsupervised and semi-supervised learning methods, incorporating principal component analysis and t-distributed stochastic neighbor embedding. To sum up, both wristbands demonstrated the usability of wearable sensors in the early detection and mitigation of falls in the elderly, outperforming the smartphone.

BACKGROUND: Negation and speculation unrelated to abnormal findings can lead to false-positive alarms for automatic radiology report highlighting or flagging by laboratory information systems.
OBJECTIVE: This internal validation study evaluated the performance of natural language processing methods (NegEx, NegBio, NegBERT, and transformers).
METHODS: We annotated all negative and speculative statements unrelated to abnormal findings in reports. In experiment 1, we fine-tuned several transformer models (ALBERT [A Lite Bidirectional Encoder Representations from Transformers], BERT [Bidirectional Encoder Representations from Transformers], DeBERTa [Decoding-Enhanced BERT With Disentangled Attention], DistilBERT [Distilled version of BERT], ELECTRA [Efficiently Learning an Encoder That Classifies Token Replacements Accurately], ERNIE [Enhanced Representation through Knowledge Integration], RoBERTa [Robustly Optimized BERT Pretraining Approach], SpanBERT, and XLNet) and compared their performance using precision, recall, accuracy, and F1-scores. In experiment 2, we compared the best model from experiment 1 with 3 established negation and speculation-detection algorithms (NegEx, NegBio, and NegBERT).
RESULTS: Our study collected 6000 radiology reports from 3 branches of the Chi Mei Hospital, covering multiple imaging modalities and body parts. A total of 15.01% (105,755/704,512) of words and 39.45% (4529/11,480) of important diagnostic keywords occurred in negative or speculative statements unrelated to abnormal findings. In experiment 1, all models achieved an accuracy of >0.98 and F1-score of >0.90 on the test data set. ALBERT exhibited the best performance (accuracy=0.991; F1-score=0.958). In experiment 2, ALBERT outperformed the optimized NegEx, NegBio, and NegBERT methods in terms of overall performance (accuracy=0.996; F1-score=0.991), in the prediction of whether diagnostic keywords occur in speculative statements unrelated to abnormal findings, and in the improvement of the performance of keyword extraction (accuracy=0.996; F1-score=0.997).
CONCLUSIONS: The ALBERT deep learning method showed the best performance. Our results represent a significant advancement in the clinical applications of computer-aided notification systems.

Exposure to harmful and potentially harmful constituents in cigarette smoke is a risk factor for cardiovascular and respiratory diseases. Tobacco products that could reduce exposure to these constituents have been developed. However, the long-term effects of their use on health remain unclear. The Population Assessment of Tobacco and Health (PATH) study is a population-based study examining the health effects of smoking and cigarette smoking habits in the U.S.
Participants include users of tobacco products, including electronic cigarettes and smokeless tobacco. In this study, we attempted to evaluate the population-wide effects of these products, using machine learning techniques and data from the PATH study.
Biomarkers of exposure (BoE) and potential harm (BoPH) in cigarette smokers and former smokers in wave 1 of PATH were used to create binary classification machine-learning models that classified participants as either current (BoE: N = 102, BoPH: N = 428) or former smokers (BoE: N = 102, BoPH: N = 428). Data on the BoE and BoPH of users of electronic cigarettes (BoE: N = 210, BoPH: N = 258) and smokeless tobacco (BoE: N = 206, BoPH: N = 242) were input into the models to investigate whether these product users were classified as current or former smokers. The disease status of individuals classified as either current or former smokers was investigated.
The classification models for BoE and BoPH both had high model accuracy. More than 60% of participants who used either one of electronic cigarettes or smokeless tobacco were classified as former smokers in the classification model for BoE. Fewer than 15% of current smokers and dual users were classified as former smokers. A similar trend was found in the classification model for BoPH. Compared with those classified as former smokers, a higher percentage of those classified as current smokers had cardiovascular disease (9.9-10.9% vs. 6.3-6.4%) and respiratory diseases (19.4-22.2% vs. 14.2-16.7%).
Users of electronic cigarettes or smokeless tobacco are likely to be similar to former smokers in their biomarkers of exposure and potential harm. This suggests that using these products helps to reduce exposure to the harmful constituents of cigarettes, and they are potentially less harmful than conventional cigarettes.

In psychology, linear discriminant analysis (LDA) is the method of choice for two-group classification tasks based on questionnaire data. In this study, we present a comparison of LDA with several supervised learning algorithms. In particular, we examine to what extent the predictive performance of LDA relies on the multivariate normality assumption. As nonparametric alternatives, the linear support vector machine (SVM), classification and regression tree (CART), random forest (RF), probabilistic neural network (PNN), and the ensemble k conditional nearest neighbor (EkCNN) algorithms are applied. Predictive performance is determined using measures of overall performance, discrimination, and calibration, and is compared in two reference data sets as well as in a simulation study. The reference data are Likert-type data, and comprise 5 and 10 predictor variables, respectively. Simulations are based on the reference data and are done for a balanced and an unbalanced scenario in each case. In order to compare the algorithms\' performance, data are simulated from multivariate distributions with differing degrees of nonnormality. Results differ depending on the specific performance measure. The main finding is that LDA is always outperformed by RF in the bimodal data with respect to overall performance. Discriminative ability of the RF algorithm is often higher compared to LDA, but its model calibration is usually worse. Still LDA mostly ranges second in cases it is outperformed by another algorithm, or the differences are only marginal. In consequence, we still recommend LDA for this type of application.

Quantifying emotional aspects of animal behavior (e.g., anxiety, social interactions, reward, and stress responses) is a major focus of neuroscience research. Because manual scoring of emotion-related behaviors is time-consuming and subjective, classical methods rely on easily quantified measures such as lever pressing or time spent in different zones of an apparatus (e.g., open vs. closed arms of an elevated plus maze). Recent advancements have made it easier to extract pose information from videos, and multiple approaches for extracting nuanced information about behavioral states from pose estimation data have been proposed. These include supervised, unsupervised, and self-supervised approaches, employing a variety of different model types. Representations of behavioral states derived from these methods can be correlated with recordings of neural activity to increase the scope of connections that can be drawn between the brain and behavior. In this mini review, we will discuss how deep learning techniques can be used in behavioral experiments and how different model architectures and training paradigms influence the type of representation that can be obtained.

BACKGROUND: Despite the numerous studies evaluating various rhythm control strategies for atrial fibrillation (AF), determination of the optimal strategy in a single patient is often based on trial and error, with no one-size-fits-all approach based on international guidelines/recommendations. The decision, therefore, remains personal and lends itself well to help from a clinical decision support system, specifically one guided by artificial intelligence (AI). QRhythm utilizes a 2-stage machine learning (ML) model to identify the optimal rhythm management strategy in a given patient based on a set of clinical factors, in which the model first uses supervised learning to predict the actions of an expert clinician and identifies the best strategy through reinforcement learning to obtain the best clinical outcome-a composite of symptomatic recurrence, hospitalization, and stroke.
OBJECTIVE: We qualitatively evaluated a novel, AI-based, clinical decision support system (CDSS) for AF rhythm management, called QRhythm, which uses both supervised and reinforcement learning to recommend either a rate control or one of 3 types of rhythm control strategies-external cardioversion, antiarrhythmic medication, or ablation-based on individual patient characteristics.
METHODS: Thirty-three clinicians, including cardiology attendings and fellows and internal medicine attendings and residents, performed an assessment of QRhythm, followed by a survey to assess relative comfort with automated CDSS in rhythm management and to examine areas for future development.
RESULTS: The 33 providers were surveyed with training levels ranging from resident to fellow to attending. Of the characteristics of the app surveyed, safety was most important to providers, with an average importance rating of 4.7 out of 5 (SD 0.72). This priority was followed by clinical integrity (a desire for the advice provided to make clinical sense; importance rating 4.5, SD 0.9), backward interpretability (transparency in the population used to create the algorithm; importance rating 4.3, SD 0.65), transparency of the algorithm (reasoning underlying the decisions made; importance rating 4.3, SD 0.88), and provider autonomy (the ability to challenge the decisions made by the model; importance rating 3.85, SD 0.83). Providers who used the app ranked the integrity of recommendations as their highest concern with ongoing clinical use of the model, followed by efficacy of the application and patient data security. Trust in the app varied; 1 (17%) provider responded that they somewhat disagreed with the statement, \"I trust the recommendations provided by the QRhythm app,\" 2 (33%) providers responded with neutrality to the statement, and 3 (50%) somewhat agreed with the statement.
CONCLUSIONS: Safety of ML applications was the highest priority of the providers surveyed, and trust of such models remains varied. Widespread clinical acceptance of ML in health care is dependent on how much providers trust the algorithms. Building this trust involves ensuring transparency and interpretability of the model.

Predicting treatment outcome in major depressive disorder (MDD) remains an essential challenge for precision psychiatry. Clinical prediction models (CPMs) based on supervised machine learning have been a promising approach for this endeavor. However, only few CPMs have focused on model sparsity even though sparser models might facilitate the translation into clinical practice and lower the expenses of their application.
In this study, we developed a predictive modeling pipeline that combines hyperparameter tuning and recursive feature elimination in a nested cross-validation framework. We applied this pipeline to a real-world clinical data set on MDD treatment response and to a second simulated data set using three different classification algorithms. Performance was evaluated by permutation testing and comparison to a reference pipeline without nested feature selection.
Across all models, the proposed pipeline led to sparser CPMs compared to the reference pipeline. Except for one comparison, the proposed pipeline resulted in equally or more accurate predictions. For MDD treatment response, balanced accuracy scores ranged between 61 and 71% when models were applied to hold-out validation data.
The resulting models might be particularly interesting for clinical applications as they could reduce expenses for clinical institutions and stress for patients.