Machine Learning in Healthcare

The paper models co-occurrence patterns of obstetric and non-obstetric diagnoses during pregnancy using graph neural networks and identifies clinically relevant comorbidity associations.

AgeML is an open-source software for age prediction from clinical data, enabling standardization, reproducibility, and analysis of age-related relationships in health and disease.

This study develops machine learning models using EHR data to predict physiological effects of multiple drugs, achieving an average AUROC of 0.75.

This paper uses transfer learning to improve disease prediction accuracy for underrepresented populations without harming accuracy for others.

The paper introduces a phenotype embedding technique using 53 million patient notes to enhance representation learning, showing improved phenotypic similarity and alignment with domain experts.

MixEHR-Guided automatically identifies interpretable phenotypes from EHR data using a multimodal topic model, outperforming existing methods without gold-standard labels.

This paper improves time-to-event prediction by integrating compatible related clinical outcomes, enhancing performance without requiring auxiliary data during prediction.

The paper introduces a graph-embedded topic model to analyze UK Biobank data, identifying pain phenotypes and revealing cardiovascular-related medication and condition patterns.

The paper presents a modular pipeline for federated health data analysis using deep learning to predict health outcomes across diseases.

SCORNET estimates survival curves using EHR data with noisy event times and limited current status labels, achieving efficiency similar to Weibull regression with non-parametric flexibility.

The paper develops a model to identify patients likely to have a genetic disease using EHR data, achieving high accuracy in predicting genetic testing needs.

The paper enhances a COVID-19 dataset by adding weather and research sentiment features, achieving 95% accuracy in survival prediction using XGBoost.

This paper develops and compares machine learning models to predict 30-day unplanned readmission in frail elderly patients using electronic health records.

The paper compares machine learning algorithms for identifying acute ischemic stroke patients using EHR data, finding high accuracy with logistic regression and demonstrating generalizability.

This study develops a high-throughput phenotyping method using temporal sequences from EHR data, achieving superior classification performance compared to conventional methods.

This paper develops indication embeddings to reduce health data dimensionality, improving cohort matching for drug safety studies by capturing relevant treatment decision information.

sureLDA automates phenotype labeling in EHR data using topic modeling, achieving high accuracy and robustness across various phenotypes.

This study compares prediction models for 30-day hospital readmissions, finding that artificial neural networks with medical code embeddings improve accuracy compared to logistic regression.

The paper develops and validates phenotype classifiers across multiple sites using machine learning, showing good portability within the US but limited international generalizability.

This paper validates a machine learning algorithm for predicting 48-hour all-cause mortality across multiple centers, showing high AUROC values compared to qSOFA and MEWS.

The paper integrates biomedical research with electronic health records to create biologically meaningful machine-readable embeddings for precision medicine.

This paper uses LSTM-based RNNs to predict ICU readmissions from EHR data, achieving higher sensitivity and AUC than traditional methods.

This study shows that natural language processing applied to electronic health records can accurately identify patients with non-alcoholic fatty liver disease and detect those at risk for disease progression.

This paper develops an automated high-throughput phenotyping method (MAP) that improves phenotype accuracy and scalability for PheWAS studies.

The paper models caregiver networks around individual patients to predict clinical outcomes like hospital stay and mortality.

Deep learning rivals or beats baselines across imaging, health records, genomics and drug discovery but is not yet transformative; scarce labels, privacy and opacity remain hurdles.

The paper develops a topic extraction pipeline from psychiatric EHRs to support readmission risk prediction, showing initial results and future feature plans.

This paper develops a composite mixture model to analyze sepsis EMR data, identifying mortality risk trends and subgroups.

The paper learns patient trajectories from EHR data using deep learning methods and compares them to traditional machine learning approaches.