Machine learning model for early prediction of acute kidney injury (AKI) in pediatric critical care.

BACKGROUND: Acute kidney injury (AKI) in pediatric critical care patients is diagnosed using elevated serum creatinine, which occurs only after kidney impairment. There are no treatments other than supportive care for AKI once it has developed, so it is important to identify patients at risk to prevent injury. This study develops a machine learning model to learn pre-disease patterns of physiological measurements and predict pediatric AKI up to 48 h earlier than the currently established diagnostic guidelines. METHODS: EHR data from 16,863 pediatric critical care patients between 1 month to 21 years of age from three independent institutions were used to develop a single machine learning model for early prediction of creatinine-based AKI using intelligently engineered predictors, such as creatinine rate of change, to automatically assess real-time AKI risk. The primary outcome is prediction of moderate to severe AKI (Stage 2/3), and secondary outcomes are prediction of any AKI (Stage 1/2/3) and requirement of renal replacement therapy (RRT). Predictions generate alerts allowing fast assessment and reduction of AKI risk, such as: "patient has 90% risk of developing AKI in the next 48 h" along with contextual information and suggested response such as "patient on aminoglycosides, suggest check level and review dose and indication". RESULTS: The model was successful in predicting Stage 2/3 AKI prior to detection by conventional criteria with a median lead-time of 30 h at AUROC of 0.89. The model predicted 70% of subsequent RRT episodes, 58% of Stage 2/3 episodes, and 41% of any AKI episodes. The ratio of false to true alerts of any AKI episodes was approximately one-to-one (PPV 47%). Among patients predicted, 79% received potentially nephrotoxic medication after being identified by the model but before development of AKI. CONCLUSIONS: As the first multi-center validated AKI prediction model for all pediatric critical care patients, the machine learning model described in this study accurately predicts moderate to severe AKI up to 48 h in advance of AKI onset. The model may improve outcome of pediatric AKI by providing early alerting and actionable feedback, potentially preventing or reducing AKI by implementing early measures such as medication adjustment.

Early prediction of hemodynamic interventions in the intensive care unit using machine learning.

BACKGROUND: Timely recognition of hemodynamic instability in critically ill patients enables increased vigilance and early treatment opportunities. We develop the Hemodynamic Stability Index (HSI), which highlights situational awareness of possible hemodynamic instability occurring at the bedside and to prompt assessment for potential hemodynamic interventions. METHODS: We used an ensemble of decision trees to obtain a real-time risk score that predicts the initiation of hemodynamic interventions an hour into the future. We developed the model using the eICU Research Institute (eRI) database, based on adult ICU admissions from 2012 to 2016. A total of 208,375 ICU stays met the inclusion criteria, with 32,896 patients (prevalence = 18%) experiencing at least one instability event where they received one of the interventions during their stay. Predictors included vital signs, laboratory measurements, and ventilation settings. RESULTS: HSI showed significantly better performance compared to single parameters like systolic blood pressure and shock index (heart rate/systolic blood pressure) and showed good generalization across patient subgroups. HSI AUC was 0.82 and predicted 52% of all hemodynamic interventions with a lead time of 1-h with a specificity of 92%. In addition to predicting future hemodynamic interventions, our model provides confidence intervals and a ranked list of clinical features that contribute to each prediction. Importantly, HSI can use a sparse set of physiologic variables and abstains from making a prediction when the confidence is below an acceptable threshold. CONCLUSIONS: The HSI algorithm provides a single score that summarizes hemodynamic status in real time using multiple physiologic parameters in patient monitors and electronic medical records (EMR). Importantly, HSI is designed for real-world deployment, demonstrating generalizability, strong performance under different data availability conditions, and providing model explanation in the form of feature importance and prediction confidence.

Association of systolic, diastolic, mean, and pulse pressure with morbidity and mortality in septic ICU patients: a nationwide observational study.

BACKGROUND: Intensivists target different blood pressure component values to manage intensive care unit (ICU) patients with sepsis. We aimed to evaluate the relationship between individual blood pressure components and organ dysfunction in critically ill septic patients. METHODS: In this retrospective observational study, we evaluated 77,328 septic patients in 364 ICUs in the eICU Research Institute database. Primary exposure was the lowest cumulative value of each component; mean, systolic, diastolic, and pulse pressure, sustained for at least 120 min during ICU stay. Primary outcome was ICU mortality and secondary outcomes were composite outcomes of acute kidney injury or death and myocardial injury or death during ICU stay. Multivariable logistic regression spline and threshold regression adjusting for potential confounders were conducted to evaluate associations between exposures and outcomes. Sensitivity analysis was conducted in 4211 patients with septic shock. RESULTS: Lower values of all blood pressures components were associated with a higher risk of ICU mortality. Estimated change-points for the risk of ICU mortality were 69 mmHg for mean, 100 mmHg for systolic, 60 mmHg for diastolic, and 57 mmHg for pulse pressure. The strength of association between blood pressure components and ICU mortality as determined by slopes of threshold regression were mean (- 0.13), systolic (- 0.11), diastolic (- 0.09), and pulse pressure (- 0.05). Equivalent non-linear associations between blood pressure components and ICU mortality were confirmed in septic shock patients. We observed a similar relationship between blood pressure components and secondary outcomes. CONCLUSION: Blood pressure component association with ICU mortality is the strongest for mean followed by systolic, diastolic, and weakest for pulse pressure. Critical care teams should continue to follow MAP-based resuscitation, though exploratory analysis focusing on blood pressure components in different sepsis phenotypes in critically ill ICU patients is needed.

Continuous noninvasive blood gas estimation in critically ill pediatric patients with respiratory failure.

Patients supported by mechanical ventilation require frequent invasive blood gas samples to monitor and adjust the level of support. We developed a transparent and novel blood gas estimation model to provide continuous monitoring of blood pH and arterial CO2 in between gaps of blood draws, using only readily available noninvasive data sources in ventilated patients. The model was trained on a derivation dataset (1,883 patients, 12,344 samples) from a tertiary pediatric intensive care center, and tested on a validation dataset (286 patients, 4030 samples) from the same center obtained at a later time. The model uses pairwise non-linear interactions between predictors and provides point-estimates of blood gas pH and arterial CO2 along with a range of prediction uncertainty. The model predicted within Clinical Laboratory Improvement Amendments of 1988 (CLIA) acceptable blood gas machine equivalent in 74% of pH samples and 80% of PCO2 samples. Prediction uncertainty from the model improved estimation accuracy by 15% by identifying and abstaining on a minority of high-uncertainty samples. The proposed model estimates blood gas pH and CO2 accurately in a large percentage of samples. The model's abstention recommendation coupled with ranked display of top predictors for each estimation lends itself to real-time monitoring of gaps between blood draws, and the model may help users determine when a new blood draw is required and delay blood draws when not needed.

Early Prediction of Cardiogenic Shock Using Machine Learning.

Cardiogenic shock (CS) is a severe condition with in-hospital mortality of up to 50%. Patients who develop CS may have previous cardiac history, but that may not always be the case, adding to the challenges in optimally identifying and managing these patients. Patients may present to a medical facility with CS or develop CS while in the emergency department (ED), in a general inpatient ward (WARD) or in the critical care unit (CC). While different clinical pathways for management exist once CS is recognized, there are challenges in identifying the patients in a timely manner, in all settings, in a timeframe that will allow proper management. We therefore developed and evaluated retrospectively a machine learning model based on the XGBoost (XGB) algorithm which runs automatically on patient data from the electronic health record (EHR). The algorithm was trained on 8 years of de-identified data (from 2010 to 2017) collected from a large regional healthcare system. The input variables include demographics, vital signs, laboratory values, some orders, and specific pre-existing diagnoses. The model was designed to make predictions 2 h prior to the need of first CS intervention (inotrope, vasopressor, or mechanical circulatory support). The algorithm achieves an overall area under curve (AUC) of 0.87 (0.81 in CC, 0.84 in ED, 0.97 in WARD), which is considered useful for clinical use. The algorithm can be refined based on specific elements defining patient subpopulations, for example presence of acute myocardial infarction (AMI) or congestive heart failure (CHF), further increasing its precision when a patient has these conditions. The top-contributing risk factors learned by the model are consistent with existing clinical findings. Our conclusion is that a useful machine learning model can be used to predict the development of CS. This manuscript describes the main steps of the development process and our results.

The Association Between Ventilatory Ratio and Mortality in Children and Young Adults.

BACKGROUND: The ventilatory ratio (VR) is a dead-space marker associated with mortality in mechanically ventilated adults with ARDS. The end-tidal alveolar dead space fraction (AVDSf) has been associated with mortality in children. However, AVDSf requires capnography measurements, whereas VR does not. We sought to examine the prognostic value of VR, in comparison to AVDSf, in children and young adults with acute hypoxemic respiratory failure. METHODS: We conducted a retrospective study of prospectively collected data from 180 mechanically ventilated children and young adults with acute hypoxemic respiratory failure. VR was calculated as (minute ventilation × [Formula: see text])/(age-adjusted predicted minute ventilation × 37.5). AVDSf was calculated as [Formula: see text]. RESULTS: VR and AVDSf had a moderate correlation (rho 0.31, P < .001). VR was similar between survivors at 1.22 (interquartile range [IQR] 1.0-1.52) and nonsurvivors at 1.30 (IQR 0.96-1.95) (P = .2). AVDSf was lower in survivors at 0.12 (IQR 0.03-0.23) than nonsurvivors at 0.24 (IQR 0.13-0.33) (P < .001). In logistic regression and competing risk regression analyses, VR was not associated with mortality or rate of extubation at any given time (competing risk death; all P > .3). An AVDSf in the highest 2 quartiles, in comparison to the lowest quartile (AVDSf < 0.06), was associated with higher mortality after adjustment for oxygenation index and severity of illness (AVDSf ≥ 0.15-0.26: odds ratio 3.58, 95% CI 1.02-12.64, P = .047, and AVDSf ≥ 0.26: odds ratio 3.91 95% CI-1.03-14.83, P = .045). At any given time after intubation, a child with an AVDSf ≥ 0.26 was less likely to be extubated than a child with an AVDSf < 0.06, after adjustment for oxygenation index and severity of illness (AVDSf ≥ 0.26: subdistribution hazard ratio 0.55, 95% CI 0.33-0.94, P = .03). CONCLUSIONS: VR should not be used for prognostic purposes in children and young adults. AVDSf added prognostic information to the severity of oxygenation defect and overall severity of illness in children and young adults, consistent with previous research.

A Physiologically Inspired Model for Solving the Cocktail Party Problem.

At a cocktail party, we can broadly monitor the entire acoustic scene to detect important cues (e.g., our names being called, or the fire alarm going off), or selectively listen to a target sound source (e.g., a conversation partner). It has recently been observed that individual neurons in the avian field L (analog to the mammalian auditory cortex) can display broad spatial tuning to single targets and selective tuning to a target embedded in spatially distributed sound mixtures. Here, we describe a model inspired by these experimental observations and apply it to process mixtures of human speech sentences. This processing is realized in the neural spiking domain. It converts binaural acoustic inputs into cortical spike trains using a multi-stage model composed of a cochlear filter-bank, a midbrain spatial-localization network, and a cortical network. The output spike trains of the cortical network are then converted back into an acoustic waveform, using a stimulus reconstruction technique. The intelligibility of the reconstructed output is quantified using an objective measure of speech intelligibility. We apply the algorithm to single and multi-talker speech to demonstrate that the physiologically inspired algorithm is able to achieve intelligible reconstruction of an "attended" target sentence embedded in two other non-attended masker sentences. The algorithm is also robust to masker level and displays performance trends comparable to humans. The ideas from this work may help improve the performance of hearing assistive devices (e.g., hearing aids and cochlear implants), speech-recognition technology, and computational algorithms for processing natural scenes cluttered with spatially distributed acoustic objects.

An ensemble boosting model for predicting transfer to the pediatric intensive care unit.

BACKGROUND: Early deterioration indicators have the potential to alert hospital care staff in advance of adverse events, such as patients requiring an increased level of care, or the need for rapid response teams to be called. Our work focuses on the problem of predicting the transfer of pediatric patients from the general ward of a hospital to the pediatric intensive care unit. OBJECTIVES: The development of a data-driven pediatric early deterioration indicator for use by clinicians with the purpose of predicting encounters where transfer from the general ward to the PICU is likely. METHODS: Using data collected over 5.5 years from the electronic health records of two medical facilities, we develop machine learning classifiers based on adaptive boosting and gradient tree boosting. We further combine these learned classifiers into an ensemble model and compare its performance to a modified pediatric early warning score (PEWS) baseline that relies on expert defined guidelines. To gauge model generalizability, we perform an inter-facility evaluation where we train our algorithm on data from one facility and perform evaluation on a hidden test dataset from a separate facility. RESULTS: We show that improvements are witnessed over the modified PEWS baseline in accuracy (0.77 vs. 0.69), sensitivity (0.80 vs. 0.68), specificity (0.74 vs. 0.70) and AUROC (0.85 vs. 0.73). CONCLUSIONS: Data-driven, machine learning algorithms can improve PICU transfer prediction accuracy compared to expertly defined systems, such as a modified PEWS, but care must be taken in the training of such approaches to avoid inadvertently introducing bias into the outcomes of these systems.

Cortical Transformation of Spatial Processing for Solving the Cocktail Party Problem: A Computational Model(1,2,3).

In multisource, "cocktail party" sound environments, human and animal auditory systems can use spatial cues to effectively separate and follow one source of sound over competing sources. While mechanisms to extract spatial cues such as interaural time differences (ITDs) are well understood in precortical areas, how such information is reused and transformed in higher cortical regions to represent segregated sound sources is not clear. We present a computational model describing a hypothesized neural network that spans spatial cue detection areas and the cortex. This network is based on recent physiological findings that cortical neurons selectively encode target stimuli in the presence of competing maskers based on source locations (Maddox et al., 2012). We demonstrate that key features of cortical responses can be generated by the model network, which exploits spatial interactions between inputs via lateral inhibition, enabling the spatial separation of target and interfering sources while allowing monitoring of a broader acoustic space when there is no competition. We present the model network along with testable experimental paradigms as a starting point for understanding the transformation and organization of spatial information from midbrain to cortex. This network is then extended to suggest engineering solutions that may be useful for hearing-assistive devices in solving the cocktail party problem.