Derivation and Validation of Generalized Sepsis-induced Acute Respiratory Failure Phenotypes Among Critically Ill Patients: A Retrospective Study.

Background: Septic patients who develop acute respiratory failure (ARF) requiring mechanical ventilation represent a heterogenous subgroup of critically ill patients with widely variable clinical characteristics. Identifying distinct phenotypes of these patients may reveal insights about the broader heterogeneity in the clinical course of sepsis. We aimed to derive novel phenotypes of sepsis-induced ARF using observational clinical data and investigate their generalizability across multi-ICU specialties, considering multi-organ dynamics. Methods: We performed a multi-center retrospective study of ICU patients with sepsis who required mechanical ventilation for ≥24 hours. Data from two different high-volume academic hospital systems were used as a derivation set with N=3,225 medical ICU (MICU) patients and a validation set with N=848 MICU patients. For the multi-ICU validation, we utilized retrospective data from two surgical ICUs at the same hospitals (N=1,577). Clinical data from 24 hours preceding intubation was used to derive distinct phenotypes using an explainable machine learning-based clustering model interpreted by clinical experts. Results: Four distinct ARF phenotypes were identified: A (severe multi-organ dysfunction (MOD) with a high likelihood of kidney injury and heart failure), B (severe hypoxemic respiratory failure [median P/F=123]), C (mild hypoxia [median P/F=240]), and D (severe MOD with a high likelihood of hepatic injury, coagulopathy, and lactic acidosis). Patients in each phenotype showed differences in clinical course and mortality rates despite similarities in demographics and admission co-morbidities. The phenotypes were reproduced in external validation utilizing an external MICU from second hospital and SICUs from both centers. Kaplan-Meier analysis showed significant difference in 28-day mortality across the phenotypes ( p <0.01) and consistent across both centers. The phenotypes demonstrated differences in treatment effects associated with high positive end-expiratory pressure (PEEP) strategy. Conclusion: The phenotypes demonstrated unique patterns of organ injury and differences in clinical outcomes, which may help inform future research and clinical trial design for tailored management strategies.

A common data model for the standardization of intensive care unit medication features.

Objective: Common data models provide a standard means of describing data for artificial intelligence (AI) applications, but this process has never been undertaken for medications used in the intensive care unit (ICU). We sought to develop a common data model (CDM) for ICU medications to standardize the medication features needed to support future ICU AI efforts. Materials and Methods: A 9-member, multi-professional team of ICU clinicians and AI experts conducted a 5-round modified Delphi process employing conference calls, web-based communication, and electronic surveys to define the most important medication features for AI efforts. Candidate ICU medication features were generated through group discussion and then independently scored by each team member based on relevance to ICU clinical decision-making and feasibility for collection and coding. A key consideration was to ensure the final ontology both distinguished unique medications and met Findable, Accessible, Interoperable, and Reusable (FAIR) guiding principles. Results: Using a list of 889 ICU medications, the team initially generated 106 different medication features, and 71 were ranked as being core features for the CDM. Through this process, 106 medication features were assigned to 2 key feature domains: drug product-related (n = 43) and clinical practice-related (n = 63). Each feature included a standardized definition and suggested response values housed in the electronic data library. This CDM for ICU medications is available online. Conclusion: The CDM for ICU medications represents an important first step for the research community focused on exploring how AI can improve patient outcomes and will require ongoing engagement and refinement.

Parsimonious Waveform-derived Features consisting of Pulse Arrival Time and Heart Rate Variability Predicts the Onset of Septic Shock.

Sepsis is a major public health emergency and one of the leading causes of morbidity and mortality in critically ill patients. For each hour treatment is delayed, shock-related mortality increases, so early diagnosis and intervention is of utmost importance. However, earlier recognition of shock requires active monitoring, which may be delayed due to subclinical manifestations of the disease at the early phase of onset. Machine learning systems can increase timely detection of shock onset by exploiting complex interactions among continuous physiological waveforms. We use a dataset consisting of high-resolution physiological waveforms from intensive care unit (ICU) of a tertiary hospital system. We investigate the use of mean arterial blood pressure (MAP), pulse arrival time (PAT), heart rate variability (HRV), and heart rate (HR) for the early prediction of shock onset. Using only five minutes of the aforementioned vital signals from 239 ICU patients, our developed models can accurately predict septic shock onset 6 to 36 hours prior to clinical recognition with area under the receiver operating characteristic (AUROC) of 0.84 and 0.8 respectively. This work lays foundations for a robust, efficient, accurate and early prediction of septic shock onset which may help clinicians in their decision-making processes. This study introduces machine learning models that provide fast and accurate predictions of septic shock onset times up to 36 hours in advance. BP, PAT and HR dynamics can independently predict septic shock onset with a look-back period of only 5 mins.

Unsupervised machine learning analysis to identify patterns of ICU medication use for fluid overload prediction.

INTRODUCTION: Intravenous (IV) medications are a fundamental cause of fluid overload (FO) in the intensive care unit (ICU); however, the association between IV medication use (including volume), administration timing, and FO occurrence remains unclear. METHODS: This retrospective cohort study included consecutive adults admitted to an ICU ≥72 hours with available fluid balance data. FO was defined as a positive fluid balance ≥7% of admission body weight within 72 hours of ICU admission. After reviewing medication administration record (MAR) data in three-hour periods, IV medication exposure was categorized into clusters using principal component analysis (PCA) and Restricted Boltzmann Machine (RBM). Medication regimens of patients with and without FO were compared within clusters to assess for temporal clusters associated with FO using the Wilcoxon rank sum test. Exploratory analyses of the medication cluster most associated with FO for medications frequently appearing and used in the first 24 hours was conducted. RESULTS: FO occurred in 127/927 (13.7%) of the patients enrolled. Patients received a median (IQR) of 31 (13-65) discrete IV medication administrations over the 72-hour period. Across all 47,803 IV medication administrations, ten unique IV medication clusters were identified with 121-130 medications in each cluster. Among the ten clusters, cluster 7 had the greatest association with FO; the mean number of cluster 7 medications received was significantly greater in patients in the FO cohort compared to patients who did not experience FO (25.6 vs.10.9. p<0.0001). 51 of the 127 medications in cluster 7 (40.2%) appeared in > 5 separate 3-hour periods during the 72-hour study window. The most common cluster 7 medications included continuous infusions, antibiotics, and sedatives/analgesics. Addition of cluster 7 medications to a prediction model with APACHE II score and receipt of diuretics improved the ability for the model to predict fluid overload (AUROC 5.65, p =0.0004). CONCLUSIONS: Using ML approaches, a unique IV medication cluster was strongly associated with FO. Incorporation of this cluster improved the ability to predict development of fluid overload in ICU patients compared with traditional prediction models. This method may be further developed into real-time clinical applications to improve early detection of adverse outcomes.

Dysfunctional neutrophil type 1 interferon responses in preschool children with recurrent wheezing and IL-4-mediated aeroallergen sensitization.

Background: The innate mechanisms associated with viral exacerbations in preschool children with recurrent wheezing are not understood. Objective: We sought to assess differential gene expression in blood neutrophils from preschool children with recurrent wheezing, stratified by aeroallergen sensitization, at baseline and after exposure to polyinosinic:polycytidylic acid (poly(I:C)) and also to examine whether poly(I:C)-stimulated blood neutrophils influenced airway epithelial gene expression. Methods: Blood neutrophils were purified and cultured overnight with poly(I:C) and underwent next-generation sequencing with Reactome pathway analysis. Primary human small airway epithelial cells were treated with poly(I:C)-treated neutrophil culture supernatants and were analyzed for type 1 interferon gene expression with a targeted array. Symptoms and exacerbations were assessed in participants over 12 months. Results: A total of 436 genes were differently expressed in neutrophils from children with versus without aeroallergen sensitization at baseline, with significant downregulation of type 1 interferons. These type 1 interferons were significantly upregulated in sensitized children after poly(I:C) stimulation. Confirmatory experiments demonstrated similar upregulation of type 1 interferons in IL-4-treated neutrophils stimulated with poly(I:C). Poly(I:C)-treated neutrophil supernatants from children with aeroallergen sensitization also induced a type 1 interferon response in epithelial cells. Children with aeroallergen sensitization also had higher symptom scores during exacerbations, and these symptom differences persisted for 3 days after prednisolone treatment. Conclusions: Type 1 interferon responses are dysregulated in preschool children with aeroallergen sensitization, which is in turn associated with exacerbation severity. Given the importance of type 1 interferon signaling in viral resolution, additional studies of neutrophil type 1 interferon responses are needed in this population.

Validation of a screening panel for pediatric metabolic dysfunction-associated steatotic liver disease using metabolomics.

BACKGROUND: Metabolic dysfunction-associated steatotic liver disease (MASLD), formerly known as NAFLD, is the most common liver disease in children. Liver biopsy remains the gold standard for diagnosis, although more efficient screening methods are needed. We previously developed a novel NAFLD screening panel in youth using machine learning applied to high-resolution metabolomics and clinical phenotype data. Our objective was to validate this panel in a separate cohort, which consisted of a combined cross-sectional sample of 161 children with stored frozen samples (75% male, 12.8±2.6 years of age, body mass index 31.0±7.0 kg/m2, 81% with MASLD, 58% Hispanic race/ethnicity). METHODS: Clinical data were collected from all children, and high-resolution metabolomics was performed using their fasting serum samples. MASLD was assessed by MRI-proton density fat fraction or liver biopsy and cardiometabolic factors. Our previously developed panel included waist circumference, triglycerides, whole-body insulin sensitivity index, 3 amino acids, 2 phospholipids, dihydrothymine, and 2 unknowns. To improve feasibility, a simplified version without the unknowns was utilized in the present study. Since the panel was modified, the data were split into training (67%) and test (33%) sets to assess the validity of the panel. RESULTS: Our present highest-performing modified model, with 4 clinical variables and 8 metabolomics features, achieved an AUROC of 0.92, 95% sensitivity, and 80% specificity for detecting MASLD in the test set. CONCLUSIONS: Therefore, this panel has promising potential for use as a screening tool for MASLD in youth.

The Pediatric Data Science and Analytics Subgroup of the Pediatric Acute Lung Injury and Sepsis Investigators Network: Use of Supervised Machine Learning Applications in Pediatric Critical Care Medicine Research.

OBJECTIVE: Perform a scoping review of supervised machine learning in pediatric critical care to identify published applications, methodologies, and implementation frequency to inform best practices for the development, validation, and reporting of predictive models in pediatric critical care. DESIGN: Scoping review and expert opinion. SETTING: We queried CINAHL Plus with Full Text (EBSCO), Cochrane Library (Wiley), Embase (Elsevier), Ovid Medline, and PubMed for articles published between 2000 and 2022 related to machine learning concepts and pediatric critical illness. Articles were excluded if the majority of patients were adults or neonates, if unsupervised machine learning was the primary methodology, or if information related to the development, validation, and/or implementation of the model was not reported. Article selection and data extraction were performed using dual review in the Covidence tool, with discrepancies resolved by consensus. SUBJECTS: Articles reporting on the development, validation, or implementation of supervised machine learning models in the field of pediatric critical care medicine. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: Of 5075 identified studies, 141 articles were included. Studies were primarily (57%) performed at a single site. The majority took place in the United States (70%). Most were retrospective observational cohort studies. More than three-quarters of the articles were published between 2018 and 2022. The most common algorithms included logistic regression and random forest. Predicted events were most commonly death, transfer to ICU, and sepsis. Only 14% of articles reported external validation, and only a single model was implemented at publication. Reporting of validation methods, performance assessments, and implementation varied widely. Follow-up with authors suggests that implementation remains uncommon after model publication. CONCLUSIONS: Publication of supervised machine learning models to address clinical challenges in pediatric critical care medicine has increased dramatically in the last 5 years. While these approaches have the potential to benefit children with critical illness, the literature demonstrates incomplete reporting, absence of external validation, and infrequent clinical implementation.

Development and Validation of a Model for Endotracheal Intubation and Mechanical Ventilation Prediction in PICU Patients.

OBJECTIVES: To develop and externally validate an intubation prediction model for children admitted to a PICU using objective and routinely available data from the electronic medical records (EMRs). DESIGN: Retrospective observational cohort study. SETTING: Two PICUs within the same healthcare system: an academic, quaternary care center (36 beds) and a community, tertiary care center (56 beds). PATIENTS: Children younger than 18 years old admitted to a PICU between 2010 and 2022. INTERVENTIONS: None. MEASUREMENTS AND MAIN RESULTS: Clinical data was extracted from the EMR. PICU stays with at least one mechanical ventilation event (≥ 24 hr) occurring within a window of 1-7 days after hospital admission were included in the study. Of 13,208 PICU stays in the derivation PICU cohort, 1,175 (8.90%) had an intubation event. In the validation cohort, there were 1,165 of 17,841 stays (6.53%) with an intubation event. We trained a Categorical Boosting (CatBoost) model using vital signs, laboratory tests, demographic data, medications, organ dysfunction scores, and other patient characteristics to predict the need of intubation and mechanical ventilation using a 24-hour window of data within their hospital stay. We compared the CatBoost model to an extreme gradient boost, random forest, and a logistic regression model. The area under the receiving operating characteristic curve for the derivation cohort and the validation cohort was 0.88 (95% CI, 0.88-0.89) and 0.92 (95% CI, 0.91-0.92), respectively. CONCLUSIONS: We developed and externally validated an interpretable machine learning prediction model that improves on conventional clinical criteria to predict the need for intubation in children hospitalized in a PICU using information readily available in the EMR. Implementation of our model may help clinicians optimize the timing of endotracheal intubation and better allocate respiratory and nursing staff to care for mechanically ventilated children.

Improving mixed-integer temporal modeling by generating synthetic data using conditional generative adversarial networks: A case study of fluid overload prediction in the intensive care unit.

OBJECTIVE: The challenge of mixed-integer temporal data, which is particularly prominent for medication use in the critically ill, limits the performance of predictive models. The purpose of this evaluation was to pilot test integrating synthetic data within an existing dataset of complex medication data to improve machine learning model prediction of fluid overload. MATERIALS AND METHODS: This retrospective cohort study evaluated patients admitted to an ICU ≥ 72 h. Four machine learning algorithms to predict fluid overload after 48-72 h of ICU admission were developed using the original dataset. Then, two distinct synthetic data generation methodologies (synthetic minority over-sampling technique (SMOTE) and conditional tabular generative adversarial network (CTGAN)) were used to create synthetic data. Finally, a stacking ensemble technique designed to train a meta-learner was established. Models underwent training in three scenarios of varying qualities and quantities of datasets. RESULTS: Training machine learning algorithms on the combined synthetic and original dataset overall increased the performance of the predictive models compared to training on the original dataset. The highest performing model was the meta-model trained on the combined dataset with 0.83 AUROC while it managed to significantly enhance the sensitivity across different training scenarios. DISCUSSION: The integration of synthetically generated data is the first time such methods have been applied to ICU medication data and offers a promising solution to enhance the performance of machine learning models for fluid overload, which may be translated to other ICU outcomes. A meta-learner was able to make a trade-off between different performance metrics and improve the ability to identify the minority class.

Machine learning-driven identification of the gene-expression signature associated with a persistent multiple organ dysfunction trajectory in critical illness.

BACKGROUND: Multiple organ dysfunction syndrome (MODS) disproportionately drives morbidity and mortality among critically ill patients. However, we lack a comprehensive understanding of its pathobiology. Identification of genes associated with a persistent MODS trajectory may shed light on underlying biology and allow for accurate prediction of those at-risk. METHODS: Secondary analyses of publicly available gene-expression datasets. Supervised machine learning (ML) was used to identify a parsimonious set of genes associated with a persistent MODS trajectory in a training set of pediatric septic shock. We optimized model parameters and tested risk-prediction capabilities in independent validation and test datasets, respectively. We compared model performance relative to an established gene-set predictive of sepsis mortality. FINDINGS: Patients with a persistent MODS trajectory had 568 differentially expressed genes and characterized by a dysregulated innate immune response. Supervised ML identified 111 genes associated with the outcome of interest on repeated cross-validation, with an AUROC of 0.87 (95% CI: 0.85-0.88) in the training set. The optimized model, limited to 20 genes, achieved AUROCs ranging from 0.74 to 0.79 in the validation and test sets to predict those with persistent MODS, regardless of host age and cause of organ dysfunction. Our classifier demonstrated reproducibility in identifying those with persistent MODS in comparison with a published gene-set predictive of sepsis mortality. INTERPRETATION: We demonstrate the utility of supervised ML driven identification of the genes associated with persistent MODS. Pending validation in enriched cohorts with a high burden of organ dysfunction, such an approach may inform targeted delivery of interventions among at-risk patients. FUNDING: H.R.W.'s NIHR35GM126943 award supported the work detailed in this manuscript. Upon his death, the award was transferred to M.N.A. M.R.A., N.S.P, and R.K were supported by NIHR21GM151703. R.K. was supported by R01GM139967.

Derivation, validation, and transcriptomic assessment of pediatric septic shock phenotypes identified through latent profile analyses: Results from a prospective multi-center observational cohort.

Background: Sepsis poses a grave threat, especially among children, but treatments are limited due to clinical and biological heterogeneity among patients. Thus, there is an urgent need for precise subclassification of patients to guide therapeutic interventions. Methods: We used clinical, laboratory, and biomarker data from a prospective multi-center pediatric septic shock cohort to derive phenotypes using latent profile analyses. Thereafter, we trained a support vector machine model to assign phenotypes in a hold-out validation set. We tested interactions between phenotypes and common sepsis therapies on clinical outcomes and conducted transcriptomic analyses to better understand the phenotype-specific biology. Finally, we compared whether newly identified phenotypes overlapped with established gene-expression endotypes and tested the utility of an integrated subclassification scheme. Findings: Among 1,071 patients included, we identified two phenotypes which we named 'inflamed' (19.5%) and an 'uninflamed' phenotype (80.5%). The 'inflamed' phenotype had an over 4-fold risk of 28-day mortality relative to those 'uninflamed'. Transcriptomic analysis revealed overexpression of genes implicated in the innate immune response and suggested an overabundance of developing neutrophils, pro-T/NK cells, and NK cells among those 'inflamed'. There was no significant overlap between endotypes and phenotypes. However, an integrated subclassification scheme demonstrated varying survival probabilities when comparing endophenotypes. Interpretation: Our research underscores the reproducibility of latent profile analyses to identify clinical and biologically informative pediatric septic shock phenotypes with high prognostic relevance. Pending validation, an integrated subclassification scheme, reflective of the different facets of the host response, holds promise to inform targeted intervention among those critically ill.

Machine learning vs. traditional regression analysis for fluid overload prediction in the ICU.

Fluid overload, while common in the ICU and associated with serious sequelae, is hard to predict and may be influenced by ICU medication use. Machine learning (ML) approaches may offer advantages over traditional regression techniques to predict it. We compared the ability of traditional regression techniques and different ML-based modeling approaches to identify clinically meaningful fluid overload predictors. This was a retrospective, observational cohort study of adult patients admitted to an ICU ≥ 72 h between 10/1/2015 and 10/31/2020 with available fluid balance data. Models to predict fluid overload (a positive fluid balance ≥ 10% of the admission body weight) in the 48-72 h after ICU admission were created. Potential patient and medication fluid overload predictor variables (n = 28) were collected at either baseline or 24 h after ICU admission. The optimal traditional logistic regression model was created using backward selection. Supervised, classification-based ML models were trained and optimized, including a meta-modeling approach. Area under the receiver operating characteristic (AUROC), positive predictive value (PPV), and negative predictive value (NPV) were compared between the traditional and ML fluid prediction models. A total of 49 of the 391 (12.5%) patients developed fluid overload. Among the ML models, the XGBoost model had the highest performance (AUROC 0.78, PPV 0.27, NPV 0.94) for fluid overload prediction. The XGBoost model performed similarly to the final traditional logistic regression model (AUROC 0.70; PPV 0.20, NPV 0.94). Feature importance analysis revealed severity of illness scores and medication-related data were the most important predictors of fluid overload. In the context of our study, ML and traditional models appear to perform similarly to predict fluid overload in the ICU. Baseline severity of illness and ICU medication regimen complexity are important predictors of fluid overload.

ClotCatcher: a novel natural language model to accurately adjudicate venous thromboembolism from radiology reports.

INTRODUCTION: Accurate identification of venous thromboembolism (VTE) is critical to develop replicable epidemiological studies and rigorous predictions models. Traditionally, VTE studies have relied on international classification of diseases (ICD) codes which are inaccurate - leading to misclassification bias. Here, we developed ClotCatcher, a novel deep learning model that uses natural language processing to detect VTE from radiology reports. METHODS: Radiology reports to detect VTE were obtained from patients admitted to Emory University Hospital (EUH) and Grady Memorial Hospital (GMH). Data augmentation was performed using the Google PEGASUS paraphraser. This data was then used to fine-tune ClotCatcher, a novel deep learning model. ClotCatcher was validated on both the EUH dataset alone and GMH dataset alone. RESULTS: The dataset contained 1358 studies from EUH and 915 studies from GMH (n = 2273). The dataset contained 1506 ultrasound studies with 528 (35.1%) studies positive for VTE, and 767 CT studies with 91 (11.9%) positive for VTE. When validated on the EUH dataset, ClotCatcher performed best (AUC = 0.980) when trained on both EUH and GMH dataset without paraphrasing. When validated on the GMH dataset, ClotCatcher performed best (AUC = 0.995) when trained on both EUH and GMH dataset with paraphrasing. CONCLUSION: ClotCatcher, a novel deep learning model with data augmentation rapidly and accurately adjudicated the presence of VTE from radiology reports. Applying ClotCatcher to large databases would allow for rapid and accurate adjudication of incident VTE. This would reduce misclassification bias and form the foundation for future studies to estimate individual risk for patient to develop incident VTE.

Enabling Continuous Breathing-Phase Contextualization via Wearable-Based Impedance Pneumography and Lung Sounds: A Feasibility Study.

Chronic respiratory diseases affect millions and are leading causes of death in the US and worldwide. Pulmonary auscultation provides clinicians with critical respiratory health information through the study of Lung Sounds (LS) and the context of the breathing-phase and chest location in which they are measured. Existing auscultation technologies, however, do not enable the simultaneous measurement of this context, thereby potentially limiting computerized LS analysis. In this work, LS and Impedance Pneumography (IP) measurements were obtained from 10 healthy volunteers while performing normal and forced-expiratory (FE) breathing maneuvers using our wearable IP and respiratory sounds (WIRS) system. Simultaneous auscultation was performed with the Eko CORE stethoscope (EKO). The breathing-phase context was extracted from the IP signals and used to compute phase-by-phase (Inspiratory (I), expiratory (E), and their ratio (I:E)) and breath-by-breath acoustic features. Their individual and added value was then elucidated through machine learning analysis. We found that the phase-contextualized features effectively captured the underlying acoustic differences between deep and FE breaths, yielding a maximum F1 Score of 84.1 ±11.4% with the phase-by-phase features as the strongest contributors to this performance. Further, the individual phase-contextualized models outperformed the traditional breath-by-breath models in all cases. The validity of the results was demonstrated for the LS obtained with WIRS, EKO, and their combination. These results suggest that incorporating breathing-phase context may enhance computerized LS analysis. Hence, multimodal sensing systems that enable this, such as WIRS, have the potential to advance LS clinical utility beyond traditional manual auscultation and improve patient care.

A MACHINE LEARNING MODEL DERIVED FROM ANALYSIS OF TIME-COURSE GENE-EXPRESSION DATASETS REVEALS TEMPORALLY STABLE GENE MARKERS PREDICTIVE OF SEPSIS MORTALITY.

ABSTRACT: Sepsis is associated with significant mortality and morbidity among critically ill patients admitted to intensive care units and represents a major health challenge globally. Given the significant clinical and biological heterogeneity among patients and the dynamic nature of the host immune response, identifying those at high risk of poor outcomes remains a critical challenge. Here, we performed secondary analysis of publicly available time-series gene-expression datasets from peripheral blood of patients admitted to the intensive care unit to elucidate temporally stable gene-expression markers between sepsis survivors and nonsurvivors. Using a limited set of genes that were determined to be temporally stable, we derived a dynamical model using a Support Vector Machine classifier to accurately predict the mortality of sepsis patients. Our model had robust performance in a test dataset, where patients' transcriptome was sampled at alternate time points, with an area under the curve of 0.89 (95% CI, 0.82-0.96) upon 5-fold cross-validation. We also identified 7 potential biomarkers of sepsis mortality (STAT5A, CX3CR1, LCP1, SNRPG, RPS27L, LSM5, SHCBP1) that require future validation. Pending prospective testing, our model may be used to identify sepsis patients with high risk of mortality accounting for the dynamic nature of the disease and with potential therapeutic implications.

Cluster analysis driven by unsupervised latent feature learning of medications to identify novel pharmacophenotypes of critically ill patients.

Unsupervised clustering of intensive care unit (ICU) medications may identify unique medication clusters (i.e., pharmacophenotypes) in critically ill adults. We performed an unsupervised analysis with Restricted Boltzmann Machine of 991 medications profiles of patients managed in the ICU to explore pharmacophenotypes that correlated with ICU complications (e.g., mechanical ventilation) and patient-centered outcomes (e.g., length of stay, mortality). Six unique pharmacophenotypes were observed, with unique medication profiles and clinically relevant differences in ICU complications and patient-centered outcomes. While pharmacophenotypes 2 and 4 had no statistically significant difference in ICU length of stay, duration of mechanical ventilation, or duration of vasopressor use, their mortality differed significantly (9.0% vs. 21.9%, p < 0.0001). Pharmacophenotype 4 had a mortality rate of 21.9%, compared with the rest of the pharmacophenotypes ranging from 2.5 to 9%. Phenotyping approaches have shown promise in classifying the heterogenous syndromes of critical illness to predict treatment response and guide clinical decision support systems but have never included comprehensive medication information. This first-ever machine learning approach revealed differences among empirically-derived subgroups of ICU patients that are not typically revealed by traditional classifiers. Identification of pharmacophenotypes may enable enhanced decision making to optimize treatment decisions.

A machine learning model for predicting congenital heart defects from administrative data.

INTRODUCTION: International Classification of Diseases (ICD) codes recorded in administrative data are often used to identify congenital heart defects (CHD). However, these codes may inaccurately identify true positive (TP) CHD individuals. CHD surveillance could be strengthened by accurate CHD identification in administrative records using machine learning (ML) algorithms. METHODS: To identify features relevant to accurate CHD identification, traditional ML models were applied to a validated dataset of 779 patients; encounter level data, including ICD-9-CM and CPT codes, from 2011 to 2013 at four US sites were utilized. Five-fold cross-validation determined overlapping important features that best predicted TP CHD individuals. Median values and 95% confidence intervals (CIs) of area under the receiver operating curve, positive predictive value (PPV), negative predictive value, sensitivity, specificity, and F1-score were compared across four ML models: Logistic Regression, Gaussian Naive Bayes, Random Forest, and eXtreme Gradient Boosting (XGBoost). RESULTS: Baseline PPV was 76.5% from expert clinician validation of ICD-9-CM CHD-related codes. Feature selection for ML decreased 7138 features to 10 that best predicted TP CHD cases. During training and testing, XGBoost performed the best in median accuracy (F1-score) and PPV, 0.84 (95% CI: 0.76, 0.91) and 0.94 (95% CI: 0.91, 0.96), respectively. When applied to the entire dataset, XGBoost revealed a median PPV of 0.94 (95% CI: 0.94, 0.95). CONCLUSIONS: Applying ML algorithms improved the accuracy of identifying TP CHD cases in comparison to ICD codes alone. Use of this technique to identify CHD cases would improve generalizability of results obtained from large datasets to the CHD patient population, enhancing public health surveillance efforts.