This is the presentation I prepared for my PhD viva.
Thesis Title: Evaluating the impact of home telehealth monitoring and clinical predictive models for patients with chronic heart failure
1. Evaluating the impact of home telehealth
monitoring and clinical predictive models for
patients with chronic heart failure
JOHN STAMFORD
2. Aims
Evaluate the efficacy of HTM for patients with CHF under the care of the
cardiology department at the Castle Hill hospital
Explore the state-of-the-art predictive models for patients with Chronic Heart
Failure (CHF)
3. Motivation
Social and Economical Impact
High number of people suffering
High cost of care
Expected to rise
HTM Efficacy
Clinical Interest
Commercial Interest
Political Interest
Background
4. Approach
Data Science
Data Extraction
Statistics
Machine Learning
Evaluation
Statistical Analysis
Cohort Comparison
Survival Analysis
Time-to-first-event
Repeat Event Analysis
Risk Model Development
Statistical Models
Machine Learning Models
Explore
Simulated Data
Clinical Data
Univariate
Bivariate
Multivariate
5. Challenges
Real Clinical Data
Missing Values
Highly Skewed
Extreme Values
Measuring HTM efficacy
No Randomised Control Trial
First Event Analysis
Need a more holistic picture
Risk Models
Definition of risk
Reported metrics
Clinical
Computer Science
Reliance on explanatory models
6. Dataset
Clinical Dataset
Cardiology Department at Castle Hill Hospital
Longitudinal
Unique
Location
Data Contained
15. Sensitivity vs Specificity
Clinical Decision
Sensitivity
Proportion of high risk patients
who are identified as high risk
Specificity
Proportion of low risk patients who
are identified as high risk
Therefore
Decision/Outcome Specific
16. Contributions
Aim 1 - Evaluate the efficacy of HTM for patients with CHF under the care of the cardiology
department at the Castle Hill hospital
Efficacy can be assessed using a longitudinal dataset
PSM can effectively match patients (reduce differences in baseline characteristics)
But – need to ensure the algorithm is ‘invariant under permutation’
HTM efficacy
Patients have better survival rates but more hospitalisation events
Spend more time alive and out of hospital
17. Contributions
Aim 2 - Explore the state-of-the-art predictive models for patients with CHF
Relationship between independent and dependant variables is not always linear / monotonic
Logistic Regression can perform similar to MLPs, and better than naïve Bayes
each model will derived different decision boundaries
Evaluation Methods
Precision, Recall, Sensitivity and Specificity are threshold dependent
Comparison – Area under the Curve
18. Future Work
Propensity score matching
Test with other datasets
Home Telehealth Monitoring
which functions of the HTM equipment, or support package, facilitated these improved
outcomes?
Can we incorporate risk into the evaluation of HTM
Need larger numbers of patients
Variable selection
Unstructured Data (MIMIC Dataset)
Editor's Notes
900,000 people in England and Wales
UK is estimated at £563 million per year
Efficacy is still unknown
WSD - At least three million people with Long Term Conditions and/or social care needs could benefit from using telehealth and telecare
The
early indications show that if used correctly telehealth can deliver a 15% reduction in A&E visits, a 20% reduction in emergency admissions, a 14% reduction in elective admissions, a 14% reduction in bed days and an 8% reduction in tariff costs. More strikingly they also demonstrate a 45% reduction in mortality rates.
6191 patients and 238 GP
External Review - the study found less evidence that telehealth will reduce costs (was finding but not statistically significant)
Political
Care Act 2014
Government Digital Strategy (2013)
Department of Health’s Digital Strategy: Leading the Culture Change in Health and Care (2012)
Department of Health’s Power of Information (2012).