European Statistical Forum

The volume of rich, real world healthcare data is growing, with the World Economic Forum estimating hospitals alone create 50 petabytes of year. Machine learning has the potential to use this data to transform aspects of how healthcare and medicine are delivered, including creating tools that facilitate personalized care and precision medicine. However, machine learning-based technologies also have the capacity to exacerbate existing inequalities or introduce new ones, and it is essential that we take steps to ensure our innovations promote health equity.

This session will highlight several case studies of how machine learning models trained on real world data are being used for precision medicine. It will also provide an overview of frameworks for understanding algorithmic fairness, explore issues specific to real-world data, and share concrete steps for creating fairer algorithms, as part of our larger goal of promoting equity in health systems. Special attention will be paid to fairness and bias considerations for large language models (e.g., ChatGPT/GPT-4, Bard/Lambda).

Co-authors:
Anel Mahmutovic, Early Biometrics & Statistical Innovation, Data Science & Artificial Intelligence, R&D, AstraZeneca, Gothenburg, Sweden
Beatriz Seoane Nuñez, Early Biometrics & Statistical Innovation, Data Science & Artificial Intelligence, R&D, AstraZeneca, Barcelona, Spain
Sofia Tapani, Early Biometrics & Statistical Innovation, Data Science & Artificial Intelligence, R&D, AstraZeneca, Gothenburg, Sweden

The term biomarker being widely used, one tends to think that its meaning is well-understood. But it is not always the case. What is a biomarker actually?
This session will (re-)introduce biomarkers and will showcase an early phase case study.

In early phase trials, the number of scientific questions that can be addressed is limited by resource constraints. Such questions could be about treatment efficacy, superiority vs. the competition, dosefinding or biomarker threshold fine-tuning.

The development of a compound initially started as a PhIIa. As interest grew, the study design became a seamless PhII, with a composite time-to-event primary endpoint.

In this study, the biomarker is continuous. Finding the biomarker threshold (BT) that will identify patients most likely to benefit from the new compound is important. To define a BT, one can start with a provisional BT and then optimize it. For BT optimization, one needs patients with biomarker values above and below the BT (respectively BM-high and BM-low).

Simulations were used to assess the number of patients needed for BT optimization. Using a stepfunction, having as many BM-low as there are BM-high patients in each treatment arm is sufficient to optimize the BT within a range of clinically meaningful values. The study was designed, approved internally and by the FDA with this number of patients.

While the study was ongoing, recruitment rates did not meet our expectations. We re-designed the study and will be able to answer most of the initial scientific questions, including the biomarker threshold optimization!

Contributors:
Oliver Pain (1), Kevin Ray(1), Karim Malki(1), Eva Krapohl(1)
1. Predictive Analytics, Statistical Sciences and Innovation, UCB Pharma, Slough, UK

Disease heterogeneity and interpatient variability contribute to differences in drug efficacy and safety – making drug development risky and costly. These individual differences are partly explained by genetic variation. Polygenic scoring is an approach for calculating an individual’s likelihood of a given outcome, leveraging data from large-scale genome-wide association studies (GWAS). Previous research supports incorporating polygenic scores into clinical trials as a biomarker to improve statistical power of the study, reduce costs, and provide novel and personalised therapeutics for patients.

Within the Predictive Analytics Team at UCB’s Statistical Sciences & Innovation department, we have developed a platform for incorporating an individual’s genetic information into the clinical trial design and analysis. The genome-wide genetic data required costs €50 − €100 per person, which once collected can simply be uploaded to our polygenic scoring platform. We then leverage the vast library of genetic associations from publicly available GWAS to calculate a range of polygenic scores that may be relevant to the outcome of interest in the clinical trial. To optimise the polygenic scores, we use Bayesian machine learning methods that capitalise on the ‘polygenic’ nature of how genetics influences disease risk. Our platform then tests whether the polygenic scores moderate the effect of the trial intervention and provides clinical and commercial insights for the design of future clinical trials.

We have applied this analysis framework to data from a clinical trial at UCB. Across all individuals, there was no effect of the drug on the clinical trial outcome, but we find evidence that polygenic scores moderate response to the drug in trial, identifying a subgroup of patients who showed significant improvements in symptoms compared to placebo.

This use case example supports the notion that incorporating polygenic scores into clinical trials can provide valuable insights into specific patient groups that will benefit from a given medication, helping to optimise clinical trials and provide personalised medicine. We present polygenic scores as a novel cost-effective biomarker for safer and more efficacious drug profiles with the potential to facilitate the drug approval process, reducing time to market and accelerating personalised medicine.

Studies in precision medicine often lead (either explicitly or implicitly) to simultaneous efficacy testing in multiple overlapping subpopulations. To limit the likelihood of false-positive conclusions, some form of control for multiple error rates is often recommended or required. Because strict control of family-wise error rate may be too conservative when the number of subpopulations is large and/or the sample sizes per subpopulation are small, less conservative approaches to multiple testing are desirable.

This presentation introduces and discusses the recently proposed concept of the population-wise error rate (PWER). The PWER controls for the average risk that a randomly selected future patient (drawn at random from the trial’s target population) will receive an ineffective treatment as a result of the trial’s test results. In this multiple error rate concept, only the type I errors relevant to the subpopulation are considered. As an average of the stratified multiple error rates, the population-wise error rate is smaller than the family-wise error rate and thus less conservative.

We will illustrate the new concept with several examples, discuss its advantages and limitations, and investigate the gain in informativeness compared to controlling for the family-wise error rate through simulation studies.

The presentation will be based on https://journals.sagepub.com/doi/full/10.1177/09622802221135249 and more recent research findings.

One of the key challenges of personalized medicine is to identify which patients will respond positively to a given treatment.

The area of subgroup identification focuses on this challenge, that is, identifying groups of patients that experience desirable characteristics, such as an enhanced treatment effect.

A crucial first step towards the subgroup identification is to identify the baseline variables (eg, biomarkers) that influence the treatment effect, which is known as predictive biomarkers. When we discover predictive biomarkers it is crucial to have control over the false-positives to avoid waste of resources, as well as provide guarantees over the replicability of our findings.

With our work we introduce a set of methods for controlled predictive biomarker discovery, and we use them to explore heterogeneity in psoriatic arthritis trials.

The promise of precision medicine is to deliver the right medicines to the right patients at the right time with the right doses.

The use of biomarkers and omics technologies provides useful tools to help realize this promise by understanding the heterogeneity in people’s genes and associating that with disease status, progression patterns, and response to treatments. For example, molecular profiling of cancers informs more precise subtypes of cancers and variation of immunophenotyping while pharmacogenomics unravels genetic associations with drug response. The interdisciplinary research of precision medicine calls for statisticians to be equipped with advanced analytical strategies and AI/ML methods while have working knowledge in clinical development and molecular biology.

In this talk, I will provide an overview of recent developments of AI/ML and omics data science applied to precision medicine across the entire lifecycle of medical product development. I will use examples of genetic risk score to demonstrate its use in drug discovery and development. I will also discuss data and statistical challenges in precision medicine and provide an outlook for the positive role statisticians can play in accelerating drug development.

One of the most crucial phases of a biomarker discovery or diagnostics development using machine learning (ML) is training the learner algorithm.

A learner is only as good as how well it has been trained in terms of biological variation which depends on the size and the heterogeneity found in the training set.

This on the other hand poses the logistical challenge of planning resources for the training phase. As far as we know, there are no well known approaches to estimating the size of the training data set.

In this talk we restrict our focus on learners where a single dominating predictor can be identified. In this we case we propose and present a simulation based approach to getting a ball-park estimate on the size of the training data.

We also present a seamless adaptive design where the training and the validation of a ML based diagnostic device can be carried out in an operationally seamless fashion while mitigating several risk factors that arise naturally in these kinds of biomedical problems.

Through an interactive tool, easily accessible from smartphones and PCs, you can take part in an engaging and informative polls on Precision Medicine and the role of Artificial Intelligence.

This is an excellent opportunity to foster collaborative discussions, encourage diverse perspectives, facilitate the exchange of ideas, and have some fun!