Add files via upload

slides/Copenhagen 2023/Copenhagen2023.Rproj

@@ -0,0 +1,13 @@
+Version: 1.0
+
+RestoreWorkspace: Default
+SaveWorkspace: Default
+AlwaysSaveHistory: Default
+
+EnableCodeIndexing: Yes
+UseSpacesForTab: Yes
+NumSpacesForTab: 2
+Encoding: UTF-8
+
+RnwWeave: Sweave
+LaTeX: pdfLaTeX

slides/Copenhagen 2023/Copenhagen2023.qmd

@@ -0,0 +1,311 @@
+---
+title: "Personalizing medical decision making"
+subtitle: "Recent advances in prediction model research"
+author: "Thomas Debray, PhD"
+format: revealjs
+engine: knitr
+---
+
+## Prediction
+
+Estimate something that is yet unknown
+
+- Presence of a certain disease (<span style="color:green;">diagnosis</span>)
+- Future occurrence of a particular event  (<span style="color:green;">prognosis</span>)
+
+![](Picture 1.png){width=100% height=400}
+
+
+## Prediction
+
+Calculate the absolute risk (probability) for distinct individuals
+
+**Why?**
+
+- Identify high-risk individuals
+- Identify absolute treatment effect
+- Target decision making to individuals
+- Causal inference
+
+![](Picture 2.png){width=100% fig-align="right"}
+
+
+## How?
+
+Calculate the absolute risk (probability) for distinct individuals
+
+**How?**
+
+Combine information from multiple predictors
+
+- Subject characteristics (e.g. age, gender)
+- History and physical examination results (e.g. blood pressure)
+- Imaging results
+- (Bio)markers
+
+  (e.g. coronary plaque)
+
+
+## Prediction
+
+Calculate the absolute risk (probability) for distinct individuals
+
+![](Picture 3.png){width=100%}
+
+
+## Prediction
+
+Develop a multivariable statistical model
+
+- Need for patient data from large cohort studies
+- Many strategies available
+
+  (Regression, decision trees, neural networks)
+
+![](Picture 4.png){width=100%}
+
+## Machine Learning
+
+- Strong focus on prediction and classification
+- Combination of data-driven algorithms
+
+    - Nearest Neighbour
+    - Recursive Partitioning
+    - Neural Network
+    - Support Vector Machine
+
+- Avoidance of modeling assumptions (e.g. additivity, linearity), resulting in high flexibility
+
+![](Picture 5.png){width=20% height=70 fig-align="right"}
+
+
+# Validation
+
+## Why do we need external validation?
+
+- The predictive performance of a model estimated on the development data is often too optimistic
+- A prognostic model should provide predictions that are valid outside the specific context of the sample that was used for model development
+- How a model was derived is of little importance if it performs well.
+
+## Causes of poor performance
+
+- Overfitting
+- Invalid predictor effects
+- Discrepancies in outcome and predictor assessment
+- Differences between study characteristics
+- Heterogeneity in case-mix variation
+
+## What is a “good” model?
+
+![](Picture 6.png){width=100%}
+
+## What is a “good” model?
+
+![](Picture 7.png){width=100%}
+
+## Current shortcomings of validation studies
+
+**Why do we need big datasets for external validation?**
+
+- External validation requires sufficient data 
+- The predictive performance of a model tends to vary across settings, populations and periods
+- Multiple external validation studies are needed to fully appreciate the generalizability of a prediction model
+
+# Meta-analysis
+
+## The rise of big data sets
+
+Data increasingly available for thousands or even millions of patients from multiple practices, hospitals, or countries.
+
+- Meta-analysis of individual participant data (IPD) from multiple studies
+
+    - Observational studies
+    - Randomized controlled trials
+
+- Analyses of databases and registry data containing e-health records
+
+![](Picture 8.png){width=100% fig-align="right"}
+
+## Individual Participant Data Meta-analyses
+
+![](Picture 9.png){width=100%}
+
+## An illustrative example
+
+**Wynants et al. previously identified >700 prediction models for coronavirus-19**
+
+We conducted a meta-analysis of participant-level data from 46 914 patients across 18 countries to externally validate the most promising models for predicting short term mortality
+
+::: {style="font-size: 0.5em"}
+- de Jong VMT, Rousset RZ, Antonio-Villa NE, Buenen AG, Van Calster B, Bello-Chavolla OY, et al. Clinical prediction models for mortality in patients with covid-19: external validation and individual participant data meta-analysis. BMJ. 2022 Jul 12;e069881.
+
+- Wynants L, Van Calster B, Bonten MMJ, Collins GS, Debray TPA, De Vos M, et al. Prediction models for diagnosis and prognosis of covid-19 infection: systematic review and critical appraisal. BMJ. 2020 Apr 7;369:BMJ Publishing Group Ltd.
+:::
+
+
+## An illustrative example
+
+![](Picture 10.png){width=100%}
+
+
+## An illustrative example
+
+![](Picture 11.png){width=100%}
+
+## Model development using IPD-MA
+
+Internal-external cross-validation
+
+![](Picture 12.png){width=100%}
+
+
+## Development and validation of ENCALS
+
+**Prognosis of amyotrophic  lateral disease**
+
+14 cohort studies (specialized ALS centres)
+
+- N = 190 to 1,936 per study (total N = 11,475)
+- Median follow-up: 97.5 months
+- Composite endpoint
+
+  (Non-invasive ventilation for more than 23h/day, or death)
+
+## Development and validation of ENCALS
+
+![](Picture 13.png){width=100%}
+
+
+## Simple versus complex modelling
+
+::: {style="font-size: 0.8em"}
+- Prediction of heart failure
+- A cohort of 871,687 individuals from 225 general practices 
+
+  (43,987 events)
+- Candidate predictors: <span style="color:green;"> *age*, *sex*, *current smoking*, *ethnicity* </span>(CE, Caucasian ethnicity), index of multiple deprivation (IMD), body mass index (BMI), creatinine level (CL), and total cholesterol (TC).
+- Implementation of internal-external cross-validation to develop simple and complex Cox regression models
+:::
+
+::: {style="font-size: 0.5em"}
+Takada T, Nijman S, Denaxas S, Snell KIE, Uijl A, Nguyen TL, et al. Internal-external cross-validation helped to evaluate the generalizability of prediction models in large clustered datasets. J Clin Epidemiol. 2021 Apr 6;137:83–91.
+:::
+
+
+## Simple versus complex modelling
+
+Development and validation of several prognostic models
+
+1. Cox regression model with four predictors (*) as linear terms
+2. Cox regression model with eight predictors; including a RCS with three knots for all continuous predictor variables, and interaction terms between all possible combinations of two variables. Estimation involves a ridge penalty term.
+
+![](Picture 14.png){width=100%}
+
+
+## Developing generalizable prediction models
+
+Stepwise estimation procedure
+
+::: {style="font-size: 0.7em"}
+- Fitting of a pre-specified GLM in each study
+- Evaluation of performance using IECV
+- Loss = f(overall performance in hold-out studies, between-study variation)
+- Expand (or reduce) model until the overall loss no longer decreases
+- Implementation in “metamisc”
+
+  https://CRAN.R-project.org/package=metamisc 
+:::
+
+
+![](Picture 15.png){width=100%}
+
+
+## Reporting guidelines
+
+![](Picture 16.png){width=100%}
+
+
+# Treatment effect modelling
+
+## Background
+
+- Causal treatment effects are estimated at the population level in randomised controlled trials (RCTs)
+- However, clinical decision is often to be made at the individual level in practice.
+- Individualized absolute treatment effects provide a natural starting point to engage in shared decision making
+
+::: {style="font-size: 0.5em"}
+Nguyen TL, Collins GS, Landais P, Le Manach Y. Counterfactual clinical prediction models could help to infer individualized treatment effects in randomized controlled trials-An illustration with the International Stroke Trial. J Clin Epidemiol. 2020 May 25;125:47–56.
+:::
+
+## Background
+
+Requirements
+
+- Move to the absolute risk scale
+- Adjust for individual patient characteristics, including 
+
+    - Prognostic variables
+
+      predicting outcome risk on reference treatment
+    - Treatment variables
+
+      with potential for effect modification
+
+- Consider counterfactual outcomes
+
+## Background
+
+An example: **The SYNTAX score II**
+
+“*The SYNTAX score II is a clinical tool that combines clinical variables with the anatomical SYNTAX score, providing expected 4-year mortality for both coronary artery bypass grafting (CABG) and percutaneous coronary intervention (PCI) — thus recommending either PCI only, CABG only or equipoise in treatment based on long-term mortality*.”
+
+::: {style="font-size: 0.5em"}
+DOI: 10.21037/acs.2018.07.02
+:::
+
+
+## Background
+
+![](Picture 17.png){width=100%}
+
+## Background
+
+![](Picture 18.png){width=100%}
+
+## Methods for treatment effect modelling
+
+![](Picture 19.png){width=100%}
+
+
+## Simulation study (1 interaction)
+
+![](Picture 20.png){width=100%}
+
+
+## Empirical example
+
+- RCT with 1:1 allocation ratio (N = 512)
+- Population: clinically diagnosed acute otitis media (AOM) in children 6 months to 5 years of age
+- Intervention: amoxicillin
+- Outcome: fever or ear pain was after 3 days’ follow-up
+- Baseline data on: treatment received, sex, presence of recurrent AOM, fever, bilateral occurrence, ear pain, presence of a runny nose, cough, tympanic membrane abnormality, and age
+
+## Empirical example
+
+![](Picture 21.png){width=100%}
+
+
+## Main findings
+
+- Small RCTs
+
+    - Hard to improve beyond risk-magnification
+    - However, the price to pay to allow for treatment-covariate interactions was small when using both shrinkage and selection, especially for the hierarchical group lasso
+
+- Large RCTs
+
+    - Shrinkage and selection still needed
+    - Allowing for all interactions was beneficial 
+
+
+

slides/Copenhagen 2023/_quarto.yml

@@ -0,0 +1,5 @@
+project:
+  title: "Copenhagen2023"
+
+editor: visual
+