Coupled Learning Enabled Stochastic Programming with Endogenous Uncertainty

Predictive analytics, empowered by machine learning, is usually followed by decision-making problems in prescriptive analytics. We extend the previous sequential prediction-optimization paradigm to a coupled scheme such that the prediction model can guide the decision problem to produce coordinated decisions yielding higher levels of performance. Specifically, for stochastic programming (SP) models with latently decision-dependent uncertainty, without any parametric assumption of the latent dependency, we develop a coupled learning enabled optimization (CLEO) algorithm in which the learning step of predicting the local dependency and the optimization step of computing a candidate decision are conducted interactively. The CLEO algorithm automatically balances the exploration and exploitation via the trust region method with active sampling. Under certain assumptions, we show that the sequence of solutions provided by CLEO converges to a directional stationary point of the original nonconvex and nonsmooth SP problem with probability 1. In addition, we present preliminary experimental results which demonstrate the computational potential of this data-driven approach.

Artificial intelligence sepsis prediction algorithm learns to say “I don’t know”

Sepsis is a leading cause of morbidity and mortality worldwide. Early identification of sepsis is important as it allows timely administration of potentially life-saving resuscitation and antimicrobial therapy. We present COMPOSER (COnformal Multidimensional Prediction Of SEpsis Risk), a deep learning model for the early prediction of sepsis, specifically designed to reduce false alarms by detecting unfamiliar patients/situations arising from erroneous data, missingness, distributional shift and data drifts. COMPOSER flags these unfamiliar cases as indeterminate rather than making spurious predictions. Six patient cohorts (515,720 patients) curated from two healthcare systems in the United States across intensive care units (ICU) and emergency departments (ED) were used to train and externally and temporally validate this model. In a sequential prediction setting, COMPOSER achieved a consistently high area under the curve (AUC) (ICU: 0.925–0.953; ED: 0.938–0.945). Out of over 6 million prediction windows roughly 20% and 8% were identified as indeterminate amongst non-septic and septic patients, respectively. COMPOSER provided early warning within a clinically actionable timeframe (ICU: 12.2 [3.2 22.8] and ED: 2.1 [0.8 4.5] hours prior to first antibiotics order) across all six cohorts, thus allowing for identification and prioritization of patients at high risk for sepsis.

Artificial Intelligence Sepsis Prediction Algorithm Learns to Say "I don't know"

Sepsis is a leading cause of morbidity and mortality worldwide. Early identification of sepsis is important as it allows timely administration of potentially life-saving resuscitation and antimicrobial therapy. We present COMPOSER (COnformal Multidimensional Prediction Of SEpsis Risk), a deep learning model for the early prediction of sepsis, specifically designed to reduce false alarms by detecting unfamiliar patients/situations arising from erroneous data, missingness, distributional shift and data drifts. COMPOSER flags these unfamiliar cases as 'indeterminate' rather than making spurious predictions. Six patient cohorts (515,720 patients) curated from two healthcare systems in the United States across intensive care units (ICU) and emergency departments (ED) were used to train and externally and temporally validate this model. In a sequential prediction setting, COMPOSER achieved a consistently high area under the curve (AUC) (ICU: 0.925-0.953; ED: 0.938-0.945). Out of over 6 million prediction windows roughly 20% and 8% were identified as 'indeterminate' amongst non-septic and septic patients, respectively. COMPOSER provided early warning within a clinically actionable timeframe (ICU: 12.2 [3.2 22.8] and ED: 2.1 [0.8 4.5] hours prior to first antibiotics order) across all six cohorts, thus allowing for identification and prioritization of patients at high risk for sepsis.

Predicting Lower Quarter Y-Balance Test Performance From Foot Characteristics

Context: The lower quarter Y-Balance Test (YBT-LQ) is associated with injury risk; however, ankle range of motion impacts YBT-LQ. Arch height and foot sensation impact static balance, but these characteristics have not yet been evaluated relative to YBT-LQ. Objective: Determine if arch height index (AHI), forefoot sensation (SEN), and ankle dorsiflexion predict YBT-LQ composite score (CS). Design: Descriptive cohort. Setting: Athletic training laboratory. Participants: Twenty general population (14 females and 6 males; mean [SD]: age 35 [18] y, weight 70.02 [16.76] kg, height 1.68 [0.12] m) participated in this study. Interventions: AHI measurement system assessed arch height in 10% (AHI10) and 90% (AHI90) weight-bearing. Two-point discrim-a-gon discs assessed sensation (SEN) at the plantar great toe, third and fifth metatarsal heads. Biplane goniometer and weight-bearing lunge tests were used to measure static and weight-bearing dorsiflexion, respectively. The YBT-LQ assessed dynamic single-leg balance. Results: For right-limb dynamic single-leg balance, AHI90 and SEN were included in the final sequential prediction equation; however, neither model significantly (P = .052 and .074) predicted variance in YBT-LQ CS. For left-limb dynamic single-leg balance, both SEN and weight-bearing lunge test were included in the final sequential prediction equation. The regression model (SEN and weight-bearing lunge test) significantly (P = .047) predicted 22% of the variance in YBT-LQ CS. Conclusions: This study demonstrates that foot characteristics may play a role in YBT-LQ CS. The authors did not assess limb dominance in this study; therefore, the authors are unable to determine which limb would be the stance versus kicking limb. However, altered SEN and weight-bearing dorsiflexion appear to be contributing factors to YBT-LQ CS.

Multiple sequential prediction errors during reward processing in the human brain

SUMMARYRecent developments in reinforcement learning, cognitive control, and systems neuroscience highlight the complimentary roles in learning of valenced reward prediction errors (RPEs) and non-valenced salience prediction errors (PEs) driven by the magnitude of surprise. A core debate in reward learning focuses on whether valenced and non-valenced PEs can be isolated in the human electroencephalogram (EEG). Here, we combine behavioral modeling and single-trial EEG regression revealing a sequence of valenced and non-valenced PEs in an interval timing task dissociating outcome valence, magnitude, and probability. Multiple regression across temporal, spatial, and frequency dimensions revealed a spatio-tempo-spectral cascade from valenced RPE value represented by the feedback related negativity event-related potential (ERP) followed by non-valenced RPE magnitude and outcome probability effects indexed by subsequent P300 and late frontal positivity ERPs. The results show that learning is supported by a sequence of multiple PEs evident in the human EEG.