There is growing interest in the area of precision medicine, which proposes that treatments can be targeted to individual patients, and often involves the use of genetic or other high-dimensional patient-level data.
In this session, organised by Dr Laura Boyle from University of Adelaide on behalf of the Medical Section, four speakers addressed some of the diverse statistical methodologies and applications in real-life studies.
The first speaker, Dr Rute Vieira from the University of Aberdeen (pictured), spoke on 'Dynamic modelling of single-case (n-of-1) data: challenges and novel applications'. This work was developed in collaboration with Professor Robin Henderson and Professor Falko Sniehotta from Newcastle University.
While statisticians might naturally worry at the idea of a study on just one patient, the data are in fact many observations of the patient over time – for instance, data from a fitness tracker, comprising many measurements of physical characteristics of a patient. The aim is not to estimate population-level, but rather patient-level parameters. These data can be modelled using time-series methodologies, such as dynamic regression models, which allow for autocorrelation and time trends.
This type of study can be useful for understanding appropriate interventions for patients, by applying different interventions to the same patient, in a crossover design with washout periods as appropriate. This was illustrated by a study of behavioural interventions to promote sun protection in patients with a rare skin disease.
Laura then temporarily left the chair to speak on work developed with Dr Lisa McFetridge (Queen’s University Belfast), Dr Özgür Asar (Acıbadem University) and Dr Jonas Wallin (Lund University), on 'Robust joint modelling: a new approach to handle time-varying outlier impacts'. The methodology she proposed was a response to the challenge of modelling haemodialysis data from a renal patient registry over time. Individual patients have measurements at different time points. The usual approach to modelling such data could be a mixed effects model allowing a different intercept and different slope over time for each patient, and assuming normally-distributed random effects. But the distributions of the data are clearly not normal.
Laura explored various robust approaches to modelling the variance, including normal mixtures, and t-distributions with time-varying degrees of freedom, in a joint model for longitudinal and survival data. A simulation study demonstrated the improvement in estimating the variance when using robust approaches over the more usual normal assumption. A question from the audience on the relevance of this study for precision medicine was answered that predictions for individual patients are more accurate, provided modelling assumptions are correct.
The third speaker was Dr David Wright from Queens University Belfast, who spoke on 'Methodological challenges for precision public health'. This work was developed in collaboration with Professor Frank Kee (Queen’s University Belfast) and Professor David Taylor-Robinson (University of Liverpool). There is an ongoing debate on the implications of an individualised approach in public health, and a concern that such an approach could widen existing inequalities in unknown ways. For instance, a focus on individual dietary risk ignores individual differences in the ability to meet the costs of a healthy diet.
Assessment of individual risk remains an issue – for example in nutritional epidemiology, where accurate dietary data remains notoriously difficult to collect, and causal explanations are limited by collinearities, confounding, and other biases. A need for causal thinking is further apparent in the targeting of appropriate public health interventions: what should be targeted, how, and when? Precision medicine methodologies such as n-of-1 studies and microrandomised trials may have a role in testing individual interventions. But unintended consequences, especially in the context of informing people about their genetic risk factors, include false positives, unnecessary tests, waste of time and resources, and worrying people. The debate on the implications of precision medicine in the public health setting is ongoing.
The final speaker, Dr Frank Dondelinger, from Lancaster University, presented work by his post-doc Farhad Hatami, in collaboration with Sach Mukherjee and Konstantinos Perrakis at the German Centre for Neurodegenerative Diseases, on 'Predicting disease progression in neurodegenerative diseases with high phenotypic variability'. He sought to address questions about differences in the rates of progression of motorneurone disease and Alzheimer’s disease, where the same biomarker can influence different diseases, but has different effects in different patients. Can we give better predictions to patients? He used a mixture model for unknown group membership, with individual trajectories over time, and penalisation methods for high dimensional genetic covariates, to predict progression of disease in different groups of patients.
His models identified optimal patient groups, and showed improvement in predictions over an approach which ignored differences between patient groups. However, he expressed a cautionary note, that while providing some extra insights, these prediction models are not a substitute for expert clinical assessment.
In conclusion, Laura thanked the speakers for their very interesting talks, highlighting the application of statistical methods in real-life examples in the rapidly expanding area of precision medicine.