Upload
institute-of-systems-science-national-university-of-singapore
View
745
Download
0
Tags:
Embed Size (px)
DESCRIPTION
Presented by Dr Liu Nan, Senior Research Scientist and Principal Investigator, Singapore General Hospital at ISS Seminar: How Analytics is Transforming Healthcare on 31 Oct 2014.
Citation preview
Medical Informatics: Computational Analytics in Healthcare Liu Nan Department of Emergency Medicine Singapore General Hospital Division of Research Health Services Research & Biostatistics Unit Singapore General Hospital
Medical Informatics: What is it?
• A discipline at the intersection of healthcare, information science, computer science, social science, and behavioural science, etc
• Deals with the resources, devices, and methods required to optimize the acquisition, storage, retrieval, and use of information in health and biomedicine
• Needs computing infrastructure, clinical guidelines, formal medical terminologies, and information and communication systems
• Application areas include nursing, clinical care, dentistry, pharmacy, public health, physical therapy, etc
2
Medical Informatics: Why Do We Need It?
• Hospitals have moved from paper-based information management to electronic health record (EHR) system. This has enabled the retrieval of massive data (e.g. free text, image, video, audio, etc)
• Computational modeling methods have been applied to a wide spectrum of applications such as big data analytics, information retrieval, robotics, bioinformatics, and medicine
• Conventional statistical and mathematical methods continue to play important roles while new emerging technologies like machine learning and data mining have established their reputations in solving complex and difficult problems
3
Medical Informatics: Research Areas
• Clinical informatics: Evaluate and refine clinical processes; Develop, implement, and refine medical decision support systems
• Public Health Informatics: Apply informatics in areas of public health, including surveillance, prevention, preparedness & health promotion
• Translational Bioinformatics: Transform biomedical data and genomic data, into proactive, predictive, preventive, and participatory health
• Other areas like bioimaging informatics, etc
4
What is Machine Learning
Machine Learning:
The Core of
Medical Informatics
5
What is Machine Learning
• Machine learning, a branch of artificial intelligence, concerns the construction and study of systems that can learn from data
6
An example of
supervised
learning: 2-class
classification
Class 1
Class 2
Decision boundary
Testing sample
Popular Machine Learning Approaches
• Decision tree learning
• Artificial neural networks
• Support vector machines
• Clustering (unsupervised learning)
• Bayesian networks
• Representation learning (feature extraction)
• Similarity and metric learning (data ranking)
7
Machine Learning vs. Biostatistics
8
Pros • Machine learning is flexible; It provides a lot of options • Machine learning usually achieves better prediction performance Cons • Some machine learning approaches are black-box systems • Predictive variables may not be statistically significant
Which one to choose? • Use traditional biostatistics for primary analysis • Use machine learning for secondary analysis • Both methods are complementary, not competing each other
What is Machine Learning
Medical Informatics Example:
Prediction of
Major Adverse Cardiac Events
(MACE)
9
Background
• Triage is the clinical process of rapidly screening large numbers of patients to assess severity and assign priority of treatment
• Currently, triage is generally done by nurses and depends on traditional vital signs and other physiological parameters
• Objective, fast and accurate risk stratification is important to quickly identify high risk patients in the Emergency Department (ED)
10
Motivation & Objective
11
• Medical resources are limited. Numbers of doctors, nurses, medical facilities may not be sufficient for fluctuating demand
• Traditional vital signs used in triage are not shown to correlate well with MACE
• To explore the utility of new variables, e.g. heart rate variability (HRV)
• To design state-of-the-art intelligent and statistical scoring methods for risk stratification in critically ill patients
Heart Rate Variability
• HRV is the beat-to-beat variation in time interval between heart beats (RR interval) under control of autonomic nervous system
• HRV has shown significant relationship between autonomic nervous system and cardiovascular mortality
• We have previously shown that HRV outperforms vital signs in risk stratification and a combined use of both performs even better
12
Study Design & Data Collection
13
Patient
Acquire ECG signals
Acquire vital signs, e.g. SpO2
Process raw ECG signal and calculate
HRV parameters
Machine learning based scoring system
Risk scores
Collected data from previous patients: • HRV parameters • Vital signs • Outcomes
Preliminary Results
14
ML MEWS TIMI
AUC 0.813 0.672 0.621
Sen. 78.9% 42.1% 78.9%
Spe. 74.1% 78.5% 36.7%
PPV 9.6% 6.4% 4.2%
NPV 99.0% 97.5% 98.0%
ML: Machine learning; MEWS: The modified early warning score; TIMI: Thrombolysis in myocardial infarction
News Report
15
Technical Challenge: Data Imbalance
16
• Some medical datasets are imbalanced where majority class is over-presented (only 5% samples of our data meet clinical outcome)
• Most machine learning techniques are not applicable with such bias dataset where majority class samples dominate decision making
• Our solution is using ensemble learning methods to manipulate data to create several balanced subsets for risk model training
• Data imbalance is common in real-world medical applications. Traditional statistical methods are usually not suitable
Technical Challenge: Variable Selection
17
• Redundant and irrelevant information may degrade prediction
performance, thus variable selection methods are needed
0
50
100
150
200
250
300
350
400
450
500
Pre
dict
or S
igni
fican
ce
Technical Challenges: System Design
18
• On the one hand, better prediction performance may require more inputs such as 12-lead ECG and vital signs, which make device big and complex
• On the other hand, light-weight and easy-to-use are the most important features for devices in ED or at home
• Need to find a trade-off between size and performance
What is Machine Learning
Medical Informatics Example:
Applications in
Other Medical Specialties
19
Pan-Asian Resuscitation Outcomes Study (PAROS)
• PAROS is a research network (10+ countries) dedicated to Pre-hospital & Emergency Care (PEC) research
• Out-of-Hospital Cardiac Arrest (OHCA) being one of the leading causes of death
• Outcome prediction using machine learning may be useful for analyzing the effects of different resuscitation strategies
• Good-quality interventions are needed
20
Retinal Vascular Abnormalities and Risk of Hypertension
• Image processing methods have been used to calculate clinical parameters
• Machine learning can be applied to investigate the association between these parameters with the risk of hypertension
21
Cheung et al. Stroke. 2013;44:2402-2408
Other Ongoing Projects
• Deriving predictors for traumatic brain injury among children (Paediatrics)
• Identification of patients at risk of walking disability 6 months post total knee arthroplasty (Physiotherapy)
• Derivation and validation of a predictive model for patients at risk of readmission (Family Medicine)
• Natural language processing and its application on unstructured medical free text for knowledge discovery (General Surgery)
• Others
22
What is Machine Learning
Medical Informatics
is Important
in Healthcare
23
Summary
24
• The aim of medical informatics is to improve patient care
• Machine learning is applicable in many different medical specialties: emergency medicine, eye, surgery, paediatrics, family medicine, allied health, etc
• Machine learning is an alternative method for data analysis
• Machine learning is complementary to statistical analysis
• Machine learning is promising when you aim for filing a patent and/or building a start-up company for commercialization