Predicting PIRA in clinical trials and observational studies

Delivered: 12 May 2023
Event: MAGNIMS
Place: Verona, Italy
Format: Conference
Source deck: 22 slides

From the slides

Deck notes

This MAGNIMS presentation examined progression independent of relapse activity across trials and observational studies. It focused on the practical choices that shape an individual prediction.

The question

Can PIRA be predicted for an individual when its definition depends on outcome measures, follow-up and the choice of time zero?

Deck outline

PIRA as a clinical and trial outcome
Evidence from clinical trials
Moving from group effects to individual prediction
Statistical framing and model groups
EDSS limitations and the definition of time zero
Intention-to-treat and decentralised approaches

Take-away

Individual PIRA prediction depends as much on a precise outcome definition and study design as it does on the modelling method.

Original deck

Slide by slide

Read the 22-slide presentation in order, or use the index to jump.

Slide 01
Predicting PIRA in clinical trials and observational studies
Open full-size slide
Arman Eshaghi, MD, PhD
NIHR Advanced Fellow
Queen Square Institute of Neurology
and
Department of Computer Science
University College London
United Kingdom
12 May 2023
Verona, Italy
Slide 02
Disclosures
Open full-size slide
Editorial Board: Neurology
Advisory board: Merck, Bristol Myers Squibb
Research grants: National Institute for Health and Care Research (NIHR), Innovate UK, Medical Research Council (MRC), Merck, Roche, Biogen.
Equity stake: Queen Square Analytics
Slide 03
Outline
Open full-size slide
- Brief literature review
- Predicting PIRA
- Towards personalized models
- Limitations
- Future: multi-center
Slide 04
PIRA as:
Open full-size slide
Outcome
Risk factor
Tur et al, JAMA Neurology, 2023. Portaccio et al, Brain, 2022.
References
- Tur et al, JAMA Neurology, 2023. Portaccio et al, Brain, 2022. — PubMed search
Slide 05
Time to event
Open full-size slide
Prognostic models
Input variables
Risk
Clinical
Outcomes
(events)
Statistical models
Blood biomarkers
Imaging
Machine learning
Genetics
Slide 06
Outcome: clinical trials
Open full-size slide
Time to 3m PIRA
RRMS on ocrelizumab
Kappos et al, JAMA Neurology, 2020
Gärtner et al, Multiple Sclerosis Journal, 2022
References
- Kappos et al, JAMA Neurology, 2020 — PubMed
- Gärtner et al, Multiple Sclerosis Journal, 2022 — PubMed
Slide 07
Clinical trials
Open full-size slide
Observational study
- ”Selected” population — Real world, diverse population
- Exclusive of comorbidities — Comorbid conditions
- Unbiased treatment effect — Bias can only be mitigated
- Short term (3 to 5 years) — Decades long
- Small to large samples (up to 3,000) — Very large samples (>10,000)
- Rich and structured data — Sparse, unstructured and missing data
- Only head to head comparison — Comparison of different treatments
Slide 08
EPIC Team et al., Annals of Neurology, 2019
Open full-size slide
References
- EPIC Team et al., Annals of Neurology, 2019 — PubMed search
Slide 09
Meier et al., JAMA Neurology, 2022
Open full-size slide
Cagol et al, JAMA Neurology, 2022
References
- Meier et al., JAMA Neurology, 2022 — PubMed search
- Cagol et al, JAMA Neurology, 2022 — PubMed
Slide 10
Can we predict PIRA for individuals?
Open full-size slide
Clinical trials
Observational studies
Subgroups
Individuals
Empirical
medicine
Stratified
Personalized
Precision
Slide 11
Statistics
Open full-size slide
Machine learning
- Explainable / (Inference) — “Black box” / (Prediction)
- 10s of variables — Millions of variables
- Variable selection / (domain expertise) — Less domain expertise
- Data efficient — Data hungry
- Group inference — Subgroup / personalized
- Linear relationship — Non-linear relationship
Slide 12
Groups
Open full-size slide
Individuals
Terminology change: from group to individual
- Coefficients — Variable importance
- Hazard ratios — Risk scores
- Data distributions — Generalizability
- P value — Accuracy, sensitivity, etc.
- Confounder — Bias
Slide 13
EDSS worsening for individuals?
Open full-size slide
Zhao et al. , Nature Digital Medicine, 2020
References
- Zhao et al. , Nature Digital Medicine, 2020 — PubMed search
Slide 14
Challenges in modeling PIRA
Open full-size slide
Slide 15
EDSS limitations
Open full-size slide
MRI-based PIRA
- Indirect: Subclinical inflammation
- Direct
- Smoldering lesions: infeasible in clinical scans
- Atrophy: Unreliable in clinical scans
Slide 16
Challenge: Defining time zero
Open full-size slide
Graf et al, Brain Communications, 2021
References
- Graf et al, Brain Communications, 2021 — PubMed
Slide 17
Challenge: timing of treatment effect
Open full-size slide
Intention to treat
vs
Per-protocol effects
Delayed treatment effects
Roos et al, Brain, 2020
References
- Roos et al, Brain, 2020 — PubMed search
Slide 18
Future: decentralized
Open full-size slide
prognostic models
Slide 19
Terrail et al. Nature Medicine, 2023
Open full-size slide
Predicting response to neoadjuvant cancer chemotherapy
*Triple-negative breast cancer (TNBC)
References
- Terrail et al. Nature Medicine, 2023 — PubMed search
Slide 20
Terrail et al. Nature Medicine, 2023
Open full-size slide
References
- Terrail et al. Nature Medicine, 2023 — PubMed search
Slide 21
Summary
Open full-size slide
Predicting PIRA observational studies
PIRA predictors/biomarkers
Limitations
Decentralized future
Slide 22
Research group website: www.ms-pinpoint.com
Open full-size slide
Contact: a.eshaghi@ucl.ac.uk