RESUMEN
This study presents Neuro-SPARK, the first scoring system developed to assess the risk of neurologic injury in pediatric and neonatal patients on extracorporeal membrane oxygenation (ECMO). Using the extracorporeal life support organization (ELSO) registry, we applied robust machine learning methodologies and clinical expertise to a 10 years dataset. We produced separate models for veno-venous (V-V ECMO) and veno-arterial (V-A ECMO) configurations due to their different risk factors and prevalence of neurologic injury. Our models identified 14 predictor variables for V-V ECMO and 20 for V-A ECMO, which demonstrated moderate accuracy in predicting neurologic injury as defined by the area under the receiver operating characteristic (AUROC) (V-V = 0.63, V-A = 0.64) and good calibration as measured by the Brier score (V-V = 0.1, V-A = 0.15). Furthermore, our post-hoc analysis identified high- and low-risk groups that may aid clinicians in targeted neuromonitoring and guide future research on ECMO-associated neurologic injury. Despite the inherent limitations, Neuro-SPARK lays the foundation for a risk-assessment tool for neurologic injury in ECMO patients, with potential implications for improved patient outcomes.
Asunto(s)
Oxigenación por Membrana Extracorpórea , Recién Nacido , Humanos , Niño , Oxigenación por Membrana Extracorpórea/efectos adversos , Oxigenación por Membrana Extracorpórea/métodos , Estudios Retrospectivos , Medición de Riesgo , Factores de Riesgo , Sistema de RegistrosRESUMEN
We consider the problem of learning with sensitive features under the privileged information setting where the goal is to learn a classifier that uses features not available (or too sensitive to collect) at test/deployment time to learn a better model at training time. We focus on tree-based learners, specifically gradient-boosted decision trees for learning with privileged information. Our methods use privileged features as knowledge to guide the algorithm when learning from fully observed (usable) features. We derive the theory, empirically validate the effectiveness of our algorithms, and verify them on standard fairness metrics.
RESUMEN
BACKGROUND: Children with congenital and acquired heart disease are at a higher risk of cardiac arrest compared to those without heart disease. Although the monitoring of cardiopulmonary resuscitation quality and extracorporeal resuscitation technologies have advanced, survival after cardiac arrest in this population has not improved. Cardiac arrest prevention, using predictive algorithms with machine learning, has the potential to reduce cardiac arrest rates. However, few studies have evaluated the use of these algorithms in predicting cardiac arrest in children with heart disease. METHODS: We collected demographic, laboratory, and vital sign information from the electronic health records (EHR) of all the patients that were admitted to a single-center pediatric cardiac intensive care unit (CICU), between 2010 and 2019, who had a cardiac arrest during their CICU admission, as well as a comparator group of randomly selected non-cardiac-arrest controls. We compared traditional logistic regression modeling against a novel adaptation of a machine learning algorithm (functional gradient boosting), using time series data to predict the risk of cardiac arrest. RESULTS: A total of 160 unique cardiac arrest events were matched to non-cardiac-arrest time periods. Using 11 different variables (vital signs and laboratory values) from the EHR, our algorithm's peak performance for the prediction of cardiac arrest was at one hour prior to the cardiac arrest (AUROC of 0.85 [0.79,0.90]), a performance that was similar to our previously published multivariable logistic regression model. CONCLUSIONS: Our novel machine learning predictive algorithm, which was developed using retrospective data that were collected from the EHR and predicted cardiac arrest in the children that were admitted to a single-center pediatric cardiac intensive care unit, demonstrated a performance that was similar to that of a traditional logistic regression model. While these results are encouraging, future research, including prospective validations with multicenter data, is warranted prior to the implementation of this algorithm as a real-time clinical decision support tool.
RESUMEN
We consider the problem of active feature elicitation in which, given some examples with all the features (say, the full Electronic Health Record), and many examples with some of the features (say, demographics), the goal is to identify the set of examples on which more information (say, lab tests) need to be collected. The observation is that some set of features may be more expensive, personal or cumbersome to collect. We propose a classifier-independent, similarity metric-independent, general active learning approach which identifies examples that are dissimilar to the ones with the full set of data and acquire the complete set of features for these examples. Motivated by four real clinical tasks, our extensive evaluation demonstrates the effectiveness of this approach. To demonstrate the generalization capabilities of the proposed approach, we consider different divergence metrics and classifiers and present consistent results across the domains.
RESUMEN
Racial and ethnic disparities in adverse pregnancy outcomes (APOs) have been well-documented in the United States, but the extent to which the disparities are present in high-risk subgroups have not been studied. To address this problem, we first applied association rule mining to the clinical data derived from the prospective nuMoM2b study cohort to identify subgroups at increased risk of developing four APOs (gestational diabetes, hypertension acquired during pregnancy, preeclampsia, and preterm birth). We then quantified racial/ethnic disparities within the cohort as well as within high-risk subgroups to assess potential effects of risk-reduction strategies. We identify significant differences in distributions of major risk factors across racial/ethnic groups and find surprising heterogeneity in APO prevalence across these populations, both in the cohort and in its high-risk subgroups. Our results suggest that risk-reducing strategies that simultaneously reduce disparities may require targeting of high-risk subgroups with considerations for the population context.
Asunto(s)
Resultado del Embarazo , Nacimiento Prematuro , Embarazo , Femenino , Recién Nacido , Humanos , Estados Unidos , Nacimiento Prematuro/epidemiología , Nacimiento Prematuro/etiología , Estudios Prospectivos , Biología Computacional , Factores de RiesgoRESUMEN
We consider the problem of modeling gestational diabetes in a clinical study and develop a domain expert-guided probabilistic model that is both interpretable and explainable. Specifically, we construct a probabilistic model based on causal independence (Noisy-Or) from a carefully chosen set of features. We validate the efficacy of the model on the clinical study and demonstrate the importance of the features and the causal independence model.
Asunto(s)
Diabetes Gestacional , Embarazo , Femenino , Humanos , Biología Computacional , Modelos Estadísticos , CausalidadRESUMEN
Importance: Polygenic risk scores (PRS) for type 2 diabetes (T2D) can improve risk prediction for gestational diabetes (GD), yet the strength of the association between genetic and lifestyle risk factors has not been quantified. Objective: To assess the association of PRS and physical activity in existing GD risk models and identify patient subgroups who may receive the most benefits from a PRS or physical activity intervention. Design, Settings, and Participants: The Nulliparous Pregnancy Outcomes Study: Monitoring Mothers-to-Be cohort was established to study individuals without previous pregnancy lasting at least 20 weeks (nulliparous) and to elucidate factors associated with adverse pregnancy outcomes. A subcohort of 3533 participants with European ancestry was used for risk assessment and performance evaluation. Participants were enrolled from October 5, 2010, to December 3, 2013, and underwent genotyping between February 19, 2019, and February 28, 2020. Data were analyzed from September 15, 2020, to November 10, 2021. Exposures: Self-reported total physical activity in early pregnancy was quantified as metabolic equivalents of task (METs). Polygenic risk scores were calculated for T2D using contributions of 84 single nucleotide variants, weighted by their association in the Diabetes Genetics Replication and Meta-analysis Consortium data. Main Outcomes and Measures: Estimation of the development of GD from clinical, genetic, and environmental variables collected in early pregnancy, assessed using measures of model discrimination. Odds ratios and positive likelihood ratios were used to evaluate the association of PRS and physical activity with GD risk. Results: A total of 3533 women were included in this analysis (mean [SD] age, 28.6 [4.9] years). In high-risk population subgroups (body mass index ≥25 or aged ≥35 years), individuals with high PRS (top 25th percentile) or low activity levels (METs <450) had increased odds of a GD diagnosis of 25% to 75%. Compared with the general population, participants with both high PRS and low activity levels had higher odds of a GD diagnosis (odds ratio, 3.4 [95% CI, 2.3-5.3]), whereas participants with low PRS and high METs had significantly reduced risk of a GD diagnosis (odds ratio, 0.5 [95% CI, 0.3-0.9]; P = .01). Conclusions and Relevance: In this cohort study, the addition of PRS was associated with the stratified risk of GD diagnosis among high-risk patient subgroups, suggesting the benefits of targeted PRS ascertainment to encourage early intervention.
Asunto(s)
Diabetes Mellitus Tipo 2 , Diabetes Gestacional , Adulto , Estudios de Cohortes , Diabetes Mellitus Tipo 2/epidemiología , Diabetes Mellitus Tipo 2/genética , Diabetes Gestacional/epidemiología , Diabetes Gestacional/genética , Ejercicio Físico , Femenino , Predisposición Genética a la Enfermedad , Humanos , EmbarazoRESUMEN
Qualitative influence statements are often provided a priori to guide learning; we answer a challenging reverse task and automatically extract them from a learned probabilistic model. We apply our Qualitative Knowledge Extraction method toward early prediction of gestational diabetes on clinical study data. Our empirical results demonstrate that the extracted rules are both interpretable and valid.
RESUMEN
We consider the problem of learning structured causal models from observational data. In this work, we use causal Bayesian networks to represent causal relationships among model variables. To this effect, we explore the use of two types of independencies-context-specific independence (CSI) and mutual independence (MI). We use CSI to identify the candidate set of causal relationships and then use MI to quantify their strengths and construct a causal model. We validate the learned models on benchmark networks and demonstrate the effectiveness when compared to some of the state-of-the-art Causal Bayesian Network Learning algorithms from observational Data.
RESUMEN
We consider the problem of learning generalized first-order representations of concepts from a small number of examples. We augment an inductive logic programming learner with 2 novel contributions. First, we define a distance measure between candidate concept representations that improves the efficiency of search for target concept and generalization. Second, we leverage richer human inputs in the form of advice to improve the sample efficiency of learning. We prove that the proposed distance measure is semantically valid and use that to derive a PAC bound. Our experiments on diverse learning tasks demonstrate both the effectiveness and efficiency of our approach.
RESUMEN
We develop a pipeline to mine complex drug interactions by combining different similarities and interaction types (molecular, structural, phenotypic, genomic etc). Our goal is to learn an optimal kernel from these heterogeneous similarities in a supervised manner. We formulate an extensible framework that can easily integrate new interaction types into a rich model. The core of our pipeline features a novel kernel-learning approach that tunes the weights of the heterogeneous similarities, and fuses them into a Similarity-based Kernel for Identifying Drug-Drug interactions and Discovery, or SKID3. Experimental evaluation on the DrugBank database shows that SKID3 effectively combines similarities generated from chemical reaction pathways (which generally improve precision) and molecular and structural fingerprints (which generally improve recall) into a single kernel that gets the best of both worlds, and consequently demonstrates the best performance.
RESUMEN
Advice-giving has been long explored in the artificial intelligence community to build robust learning algorithms when the data is noisy, incorrect or even insufficient. While logic based systems were effectively used in building expert systems, the role of the human has been restricted to being a "mere labeler" in recent times. We hypothesize and demonstrate that probabilistic logic can provide an effective and natural way for the expert to specify domain advice. Specifically, we consider different types of advice-giving in relational domains where noise could arise due to systematic errors or class-imbalance inherent in the domains. The advice is provided as logical statements or privileged features that are thenexplicitly considered by an iterative learning algorithm at every update. Our empirical evidence shows that human advice can effectively accelerate learning in noisy, structured domains where so far humans have been merely used as labelers or as designers of the (initial or final) structure of the model.
RESUMEN
Adverse drug events (ADEs) are a major concern and point of emphasis for the medical profession, government, and society. A diverse set of techniques from epidemiology, statistics, and computer science are being proposed and studied for ADE discovery from observational health data (e.g., EHR and claims data), social network data (e.g., Google and Twitter posts), and other information sources. Methodologies are needed for evaluating, quantitatively measuring, and comparing the ability of these various approaches to accurately discover ADEs. This work is motivated by the observation that text sources such as the Medline/Medinfo library provide a wealth of information on human health. Unfortunately, ADEs often result from unexpected interactions, and the connection between conditions and drugs is not explicit in these sources. Thus, in this work we address the question of whether we can quantitatively estimate relationships between drugs and conditions from the medical literature. This paper proposes and studies a state-of-the-art NLP-based extraction of ADEs from text.
RESUMEN
Machine learning is one of the most exciting and rapidly expanding fields within computer science. Academic and commercial research entities are investing in machine learning methods, especially in personalized medicine via patient-level classification. There is great promise that machine learning methods combined with resting state functional MR imaging will aid in diagnosis of disease and guide potential treatment for conditions thought to be impossible to identify based on imaging alone, such as psychiatric disorders. We discuss machine learning methods and explore recent advances.
Asunto(s)
Encefalopatías/diagnóstico por imagen , Mapeo Encefálico/métodos , Encéfalo/diagnóstico por imagen , Procesamiento de Imagen Asistido por Computador/métodos , Aprendizaje Automático , Imagen por Resonancia Magnética/métodos , Algoritmos , Encéfalo/fisiopatología , Encefalopatías/fisiopatología , Humanos , DescansoRESUMEN
Parkinson's, a progressive neural disorder, is difficult to identify due to the hidden nature of the symptoms associated. We present a machine learning approach that uses a definite set of features obtained from the Parkinsons Progression Markers Initiative(PPMI) study as input and classifies them into one of two classes: PD(Parkinson's disease) and HC(Healthy Control). As far as we know this is the first work in applying machine learning algorithms for classifying patients with Parkinson's disease with the involvement of domain expert during the feature selection process. We evaluate our approach on 1194 patients acquired from Parkinsons Progression Markers Initiative and show that it achieves a state-of-the-art performance with minimal feature engineering.
RESUMEN
Adverse drug events (ADEs) are a major concern and point of emphasis for the medical profession, government, and society in general. When methods extract ADEs from observational data, there is a necessity to evaluate these methods. More precisely, it is important to know what is already known in the literature. Consequently, we employ a novel relation extraction technique based on a recently developed probabilistic logic learning algorithm that exploits human advice. We demonstrate on a standard adverse drug events data base that the proposed approach can successfully extract existing adverse drug events from limited amount of training data and compares favorably with state-of-the-art probabilistic logic learning methods.
RESUMEN
The pharmaceutical industry, consumer protection groups, users of medications and government oversight agencies are all strongly interested in identifying adverse reactions to drugs. While a clinical trial of a drug may use only a thousand patients, once a drug is released on the market it may be taken by millions of patients. As a result, in many cases adverse drug events (ADEs) are observed in the broader population that were not identified during clinical trials. Therefore, there is a need for continued, post-marketing surveillance of drugs to identify previously-unanticipated ADEs. This paper casts this problem as a reverse machine learning task, related to relational subgroup discovery and provides an initial evaluation of this approach based on experiments with an actual EMR/EHR and known adverse drug events.
RESUMEN
Learning temporal dependencies between variables over continuous time is an important and challenging task. Continuous-time Bayesian networks effectively model such processes but are limited by the number of conditional intensity matrices, which grows exponentially in the number of parents per variable. We develop a partition-based representation using regression trees and forests whose parameter spaces grow linearly in the number of node splits. Using a multiplicative assumption we show how to update the forest likelihood in closed form, producing efficient model updates. Our results show multiplicative forests can be learned from few temporal trajectories with large gains in performance and scalability.