Scalable and accurate deep learning with electronic health NEMO English Seminar Scalable and accurate deep learning with electronic health Alvin Rajkomar et al. Google Inc., Mountain View, CA, USA; University of California, San Francisco, San Francisco, CA,USA npj Digital Medicine, 1(1):1-18, 2018. Speaker JUNGJU PARK Date 4th July 2018
Introduction Key words Limited traditional predictive modeling Disease prediction Electronic health record (EHR) Deep learning Limited traditional predictive modeling Required custom data set with specific variables Thousands of potential predictors Variable selections may produce imprecise prediction Contribution Reporting a generic data processing pipeline Demonstration in a wide variety of predictive problems
Unstructured data (raw)
Structured data (input) Fast Healthcare Interoperability Resources (FHIR)
Descriptive statistics
Results: 216,221 hospitalizations Deep learning Logistic regression
Improving Palliative Care with Deep Learning NEMO English Seminar Improving Palliative Care with Deep Learning Anand Avati et al. Department of Computer Science, Stanford University Bioinformatics and Biomedicine (BIBM), 2017 IEEE International Conference pp. 311-316. IEEE Speaker JUNGJU PARK Date 4th July 2018
Introduction Key words Difficulties in real world Goal Palliative care Deep learning Difficulties in real world Limited resource Physicians may not refer patients Shortage of palliative care professionals Goal Identification of the patients who needs palliative cares To predict the mortality of a given patient within 12 months
Literature review Prognostic tools in Palliative Care Palliative Prognostic Score (multiple regression analysis) Prognostic tools in the Intensive Care Unit APACHE-II, -III, SAPS II, etc. Prognostic tools for Early Identification CriSTAL, CARING, etc., Prognostic in the age of Big-Data Machine Learning-based approaches Small data set, few variables, naïve model,
Procedures Dataset construction Feature extraction Training Evaluation Deep Neural Network 13,654 dimensions 18 hidden layers Evaluation With imbalanced data Average Precession (AP) score
Results Brier score of 0.042 AP score of 0.69 Recall of 0.34 at 0.9 precision AUROC of 0.93