Deep Learning for Identification of Acute Illness and Facial Cues of Illness.

Background: The inclusion of facial and bodily cues (clinical gestalt) in machine learning (ML) models improves the assessment of patients' health status, as shown in genetic syndromes and acute coronary syndrome. It is unknown if the inclusion of clinical gestalt improves ML-based classification of acutely ill patients. As in previous research in ML analysis of medical images, simulated or augmented data may be used to assess the usability of clinical gestalt. Objective: To assess whether a deep learning algorithm trained on a dataset of simulated and augmented facial photographs reflecting acutely ill patients can distinguish between healthy and LPS-infused, acutely ill individuals. Methods: Photographs from twenty-six volunteers whose facial features were manipulated to resemble a state of acute illness were used to extract features of illness and generate a synthetic dataset of acutely ill photographs, using a neural transfer convolutional neural network (NT-CNN) for data augmentation. Then, four distinct CNNs were trained on different parts of the facial photographs and concatenated into one final, stacked CNN which classified individuals as healthy or acutely ill. Finally, the stacked CNN was validated in an external dataset of volunteers injected with lipopolysaccharide (LPS). Results: In the external validation set, the four individual feature models distinguished acutely ill patients with sensitivities ranging from 10.5% (95% CI, 1.3-33.1% for the skin model) to 89.4% (66.9-98.7%, for the nose model). Specificity ranged from 42.1% (20.3-66.5%) for the nose model and 94.7% (73.9-99.9%) for skin. The stacked model combining all four facial features achieved an area under the receiver characteristic operating curve (AUROC) of 0.67 (0.62-0.71) and distinguished acutely ill patients with a sensitivity of 100% (82.35-100.00%) and specificity of 42.11% (20.25-66.50%). Conclusion: A deep learning algorithm trained on a synthetic, augmented dataset of facial photographs distinguished between healthy and simulated acutely ill individuals, demonstrating that synthetically generated data can be used to develop algorithms for health conditions in which large datasets are difficult to obtain. These results support the potential of facial feature analysis algorithms to support the diagnosis of acute illness.

Identifying and characterizing high-risk clusters in a heterogeneous ICU population with deep embedded clustering.

Critically ill patients constitute a highly heterogeneous population, with seemingly distinct patients having similar outcomes, and patients with the same admission diagnosis having opposite clinical trajectories. We aimed to develop a machine learning methodology that identifies and provides better characterization of patient clusters at high risk of mortality and kidney injury. We analysed prospectively collected data including co-morbidities, clinical examination, and laboratory parameters from a minimally-selected population of 743 patients admitted to the ICU of a Dutch hospital between 2015 and 2017. We compared four clustering methodologies and trained a classifier to predict and validate cluster membership. The contribution of different variables to the predicted cluster membership was assessed using SHapley Additive exPlanations values. We found that deep embedded clustering yielded better results compared to the traditional clustering algorithms. The best cluster configuration was achieved for 6 clusters. All clusters were clinically recognizable, and differed in in-ICU, 30-day, and 90-day mortality, as well as incidence of acute kidney injury. We identified two high mortality risk clusters with at least 60%, 40%, and 30% increased. ICU, 30-day and 90-day mortality, and a low risk cluster with 25-56% lower mortality risk. This machine learning methodology combining deep embedded clustering and variable importance analysis, which we made publicly available, is a possible solution to challenges previously encountered by clustering analyses in heterogeneous patient populations and may help improve the characterization of risk groups in critical care.

Ensemble machine learning prediction and variable importance analysis of 5-year mortality after cardiac valve and CABG operations.

Despite having a similar post-operative complication profile, cardiac valve operations are associated with a higher mortality rate compared to coronary artery bypass grafting (CABG) operations. For long-term mortality, few predictors are known. In this study, we applied an ensemble machine learning (ML) algorithm to 88 routinely collected peri-operative variables to predict 5-year mortality after different types of cardiac operations. The Super Learner algorithm was trained using prospectively collected peri-operative data from 8241 patients who underwent cardiac valve, CABG and combined operations. Model performance and calibration were determined for all models, and variable importance analysis was conducted for all peri-operative parameters. Results showed that the predictive accuracy was the highest for solitary mitral (0.846 [95% CI 0.812-0.880]) and solitary aortic (0.838 [0.813-0.864]) valve operations, confirming that ensemble ML using routine data collected perioperatively can predict 5-year mortality after cardiac operations with high accuracy. Additionally, post-operative urea was identified as a novel and strong predictor of mortality for several types of operation, having a seemingly additive effect to better known risk factors such as age and postoperative creatinine.

A Bayesian Network Analysis of the Diagnostic Process and Its Accuracy to Determine How Clinicians Estimate Cardiac Function in Critically Ill Patients: Prospective Observational Cohort Study.

BACKGROUND: Hemodynamic assessment of critically ill patients is a challenging endeavor, and advanced monitoring techniques are often required to guide treatment choices. Given the technical complexity and occasional unavailability of these techniques, estimation of cardiac function based on clinical examination is valuable for critical care physicians to diagnose circulatory shock. Yet, the lack of knowledge on how to best conduct and teach the clinical examination to estimate cardiac function has reduced its accuracy to almost that of "flipping a coin." OBJECTIVE: The aim of this study was to investigate the decision-making process underlying estimates of cardiac function of patients acutely admitted to the intensive care unit (ICU) based on current standardized clinical examination using Bayesian methods. METHODS: Patient data were collected as part of the Simple Intensive Care Studies-I (SICS-I) prospective cohort study. All adult patients consecutively admitted to the ICU with an expected stay longer than 24 hours were included, for whom clinical examination was conducted and cardiac function was estimated. Using these data, first, the probabilistic dependencies between the examiners' estimates and the set of clinically measured variables upon which these rely were analyzed using a Bayesian network. Second, the accuracy of cardiac function estimates was assessed by comparison to the cardiac index values measured by critical care ultrasonography. RESULTS: A total of 1075 patients were included, of which 783 patients had validated cardiac index measurements. A Bayesian network analysis identified two clinical variables upon which cardiac function estimate is conditionally dependent, namely, noradrenaline administration and presence of delayed capillary refill time or mottling. When the patient received noradrenaline, the probability of cardiac function being estimated as reasonable or good P(ER,G) was lower, irrespective of whether the patient was mechanically ventilated (P[ER,G|ventilation, noradrenaline]=0.63, P[ER,G|ventilation, no noradrenaline]=0.91, P[ER,G|no ventilation, noradrenaline]=0.67, P[ER,G|no ventilation, no noradrenaline]=0.93). The same trend was found for capillary refill time or mottling. Sensitivity of estimating a low cardiac index was 26% and 39% and specificity was 83% and 74% for students and physicians, respectively. Positive and negative likelihood ratios were 1.53 (95% CI 1.19-1.97) and 0.87 (95% CI 0.80-0.95), respectively, overall. CONCLUSIONS: The conditional dependencies between clinical variables and the cardiac function estimates resulted in a network consistent with known physiological relations. Conditional probability queries allow for multiple clinical scenarios to be recreated, which provide insight into the possible thought process underlying the examiners' cardiac function estimates. This information can help develop interactive digital training tools for students and physicians and contribute toward the goal of further improving the diagnostic accuracy of clinical examination in ICU patients. TRIAL REGISTRATION: ClinicalTrials.gov NCT02912624; https://clinicaltrials.gov/ct2/show/NCT02912624.

Performance of neural networks for localizing moving objects with an artificial lateral line.

Fish are able to sense water flow velocities relative to their body with their mechanoreceptive lateral line organ. This organ consists of an array of flow detectors distributed along the fish body. Using the excitation of these individual detectors, fish can determine the location of nearby moving objects. Inspired by this sensory modality, it is shown here how neural networks can be used to extract an object's location from simulated excitation patterns, as can be measured along arrays of stationary artificial flow velocity sensors. The applicability, performance and robustness with respect to input noise of different neural network architectures are compared. When trained and tested under high signal to noise conditions (46 dB), the Extreme Learning Machine architecture performs best with a mean Euclidean error of 0.4% of the maximum depth of the field D, which is taken half the length of the sensor array. Under lower signal to noise conditions Echo State Networks, having recurrent connections, enhance the performance while the Multilayer Perceptron is shown to be the most noise robust architecture. Neural network performance decreased when the source moves close to the sensor array or to the sides of the array. For all considered architectures, increasing the number of detectors per array increased localization performance and robustness.

Neural-fitted TD-leaf learning for playing Othello with structured neural networks.

This paper describes a methodology for quickly learning to play games at a strong level. The methodology consists of a novel combination of three techniques, and a variety of experiments on the game of Othello demonstrates their usefulness. First, structures or topologies in neural network connectivity patterns are used to decrease the number of learning parameters and to deal more effectively with the structural credit assignment problem, which is to change individual network weights based on the obtained feedback. Furthermore, the structured neural networks are trained with the novel neural-fitted temporal difference (TD) learning algorithm to create a system that can exploit most of the training experiences and enhance learning speed and performance. Finally, we use the neural-fitted TD-leaf algorithm to learn more effectively when look-ahead search is performed by the game-playing program. Our extensive experimental study clearly indicates that the proposed method outperforms linear networks and fully connected neural networks or evaluation functions evolved with evolutionary algorithms.

How longer saccade latencies lead to a competition for salience.

It has been suggested that independent bottom-up and top-down processes govern saccadic selection. However, recent findings are hard to explain in such terms. We hypothesized that differences in visual-processing time can explain these findings, and we tested this using search displays containing two deviating elements, one requiring a short processing time and one requiring a long processing time. Following short saccade latencies, the deviation requiring less processing time was selected most frequently. This bias disappeared following long saccade latencies. Our results suggest that an element that attracts eye movements following short saccade latencies does so because it is the only element processed at that time. The temporal constraints of processing visual information therefore seem to be a determining factor in saccadic selection. Thus, relative saliency is a time-dependent phenomenon.

Saccadic selection and crowding in visual search: stronger lateral masking leads to shorter search times.

We investigated the role of crowding in saccadic selection during visual search. To guide eye movements, often information from the visual periphery is used. Crowding is known to deteriorate the quality of peripheral information. In four search experiments, we studied the role of crowding, by accompanying individual search elements by flankers. Varying the difference between target and flankers allowed us to manipulate crowding strength throughout the stimulus. We found that eye movements are biased toward areas with little crowding for conditions where a target could be discriminated peripherally. Interestingly, for conditions in which the target could not be discriminated peripherally, this bias reversed to areas with strong crowding. This led to shorter search times for a target presented in areas with stronger crowding, compared to a target presented in areas with less crowding. These findings suggest a dual role for crowding in visual search. The presence of flankers similar to the target deteriorates the quality of the peripheral target signal but can also attract eye movements, as more potential targets are present over the area.

Ensemble algorithms in reinforcement learning.

This paper describes several ensemble methods that combine multiple different reinforcement learning (RL) algorithms in a single agent. The aim is to enhance learning speed and final performance by combining the chosen actions or action probabilities of different RL algorithms. We designed and implemented four different ensemble methods combining the following five different RL algorithms: Q-learning, Sarsa, actor-critic (AC), QV-learning, and AC learning automaton. The intuitively designed ensemble methods, namely, majority voting (MV), rank voting, Boltzmann multiplication (BM), and Boltzmann addition, combine the policies derived from the value functions of the different RL algorithms, in contrast to previous work where ensemble methods have been used in RL for representing and learning a single value function. We show experiments on five maze problems of varying complexity; the first problem is simple, but the other four maze tasks are of a dynamic or partially observable nature. The results indicate that the BM and MV ensembles significantly outperform the single RL algorithms.

Recurrent neural network modeling of nearshore sandbar behavior.

The temporal evolution of nearshore sandbars (alongshore ridges of sand fringing coasts in water depths less than 10 m and of paramount importance for coastal safety) is commonly predicted using process-based models. These models are autoregressive and require offshore wave characteristics as input, properties that find their neural network equivalent in the NARX (Nonlinear AutoRegressive model with eXogenous input) architecture. Earlier literature results suggest that the evolution of sandbars depends nonlinearly on the wave forcing and that the sandbar position at a specific moment contains 'memory', that is, time-series of sandbar positions show dependencies spanning several days. Using observations of an outer sandbar collected daily for over seven years at the double-barred Surfers Paradise, Gold Coast, Australia several data-driven models are compared. Nonlinear and linear models as well as recurrent and nonrecurrent parameter estimation methods are applied to investigate the claims about nonlinear and long-term dependencies. We find a small performance increase for long-term predictions (>40 days) with nonlinear models, indicating that nonlinear effects expose themselves for larger prediction horizons, and no significant difference between nonrecurrent and recurrent methods meaning that the effects of dependencies spanning several days are of no importance.