Machine Learning in Healthcare and Biomedical Applications

• Data driven decision making
• Questions -> Data -> Models/Tools

Table of Contents

1. Overview
2. EHR data
3. Insurance claims data
4. Clinical notes
5. Image data
6. Time series data
7. Genomics data

Overview

Data type	Models/Tools	Applications
-EHR data -Insurance claims data	ML(logistic regression,XGBoost)	Predict outcomes (disease, death, readmission etc.)
-Clinical notes -Conversation text data	-Rule based approach(regular expression) -Deep learning apporach	-Extract concepts from clinical notes -Knowledge graphs -Chat-bot -QA system
Medical image data (X-ray, CT, OCR image etc.)	CNN	-Detection: diagnosis of skin cancer lung nodule or diabetic reinopathy -Segmentation of tumor, histopathology
Time series data (EEG, ECG, vital sign data etc.)	HMM,RNN,CNN	-Heart disease -Sleep disorder(apnea) -ICU monitoring
Genomics data	GATK,QIIME	-Cancer mutation identification -Biomarker identification -Durg discovery
Other data (hospital operational data)	-ML(regression) -Queueing model	-Reduce operational cost -Improve patient experience -ER wait time and queueing

EHR data

Prediction outcomes	Models/Tools	Data type	Sample size	Reference	Year
Review				Mining electronic health records: towards better research applications and clinical care	2012
Review				Opportunities and challenges in developing risk prediction models with electronic health records data: a systematic review	2016
heart failure	-logistic regression -random forest	longitudinal EHR data	1684 heart failure cases and 13525 matched controls	Early Detection of Heart Failure Using Electronic Health Records	2016
heart failure (review)				Population Risk Prediction Models for Incident Heart Failure	2015
Kidney transplant graft failure	Cox regression	10-years EHR data	69,440 kidney transpants	A comprehensive risk quantification score for deceased donor kidneys: the kidney donor risk index	2009

Clinical notes

Prediction outcomes	Models/Tools	Data type	Sample size	Reference	Year
Review				Realizing the full potential of electronic health records: the role of natural language processing	2011
Review				Natural language processing: an introduction	2011
Negation	Regular expression and rule-based approach	Clinical reports	2060 discharge summaries	A simple algorithm for identifying negated findings and diseases in discharge summaries	2001
				Using electronic health records to drive discovery in disease genomics
NER		discharge summaries	826 notes	A study of machine-learning-based approaches to extract clinical entities and their assertions from discharge summaries	2011

Image data

Prediction outcomes	Models/Tools	Data type	Sample size	Reference	Year
Diabetic retinopathy	CNN	retinal fundus images	128175 retinal images	Development and Validation of a Deep Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus Photographs	2016
Skin cancer	CNN	skin images	129,450 skin images	Dermatologist-level classification of skin cancer with deep neural networks	2017
Tumor	CNN	Pathology images	400+110 slides	Detecting Cancer Metastases on Gigapixel Pathology Images	2017
				Deep Learning Automates the Quantitative Analysis of Individual Cells in Live-Cell Imaging Experiments

Time series data

Prediction outcomes	Models/Tools	Data type	Sample size	Reference	Year
sinus rhythm and atrial fibrillation	34-layer convolutional neural network (CNN)	single-lead ECG	-(Train) 64,121 ECG records from 29,163 patients -(Test) 336 records from 328 unique patients	Cardiologist-Level Arrhythmia Detection with Convolutional Neural Networks	2017
Hand movements	CNN	sEMG	67 intact subjects and 11 transradial amputees	Deep Learning with Convolutional Neural Networks Applied to Electromyography Data: A Resource for the Classification of Movements for Prosthetic Hands	2016
Review		ICU data		Machine Learning and Decision Support in Critical Care	2017

Genomics data

Prediction outcomes	Models/Tools	Data type	Sample size	Reference	Year
Genetic variants	Exome NGS	NGS&EHR data	50,726 individuals	Distribution and clinical impact of functional variants in 50,726 whole-exome sequences from the DiscovEHR study	2016
Familial hypercholesterolemia	Exome NGS	NGS&EHR data	50,726 individuals	Genetic identification of familial hypercholesterolemia within a single U.S. health care system	2016

Other

Prediction outcomes	Models/Tools	Data type	Sample size	Reference	Year
Drug discovery	LSTM	Assay	12-27 assays	Low data drug discovery with one-shot learning	2017
Tutorial		Image		Deep learning models for health care: challenges and solutions	2017
Tutorial		Image		Deep learning in radiology: recent advances, challenges and future trends	2016
Tutorial				Big data analytics for healthcare	2013
Tutorial		Image		Survey of deep learning in radiology	2017
ER wait time		ER visit time		Accurate ED Wait Time Prediction	2017