Joint modelling of multivariate longitudinal and event time outcomes in clinical research

Progression of a disease or condition as it develops over time can be monitored in many ways, which can be collectively referred to as biomarkers.

This proposal aims to develop novel statistical methods to investigate how the patterns over time in combination of biomarkers relate to prognosis for the patient and in particular the timing of clinically significant events. For example, sepsis is a common and potentially life-threatening condition triggered by an infection, and it emerges as a leading cause of death with a mortality rate of 30 to 50%. Hence, early detection and timely therapeutic intervention is crucial for improved outcome of patients with sepsis. However, the diagnosis of sepsis is extremely difficult due to the highly variable symptoms of sepsis, and biomarkers hold the promise to separate sepsis from local infection. It is unlikely that a single ideal biomarker will ever be found in view of the complexity of the sepsis response, but a combination of biomarkers may be more effective for real-time indication and prognostic information of sepsis. Hence, it is not surprising that in most medical disciplines including sepsis, the current research is focused on patterns of biomarkers combinations (rather than a single biomarker) over time in relation to the timing of clinical events.

Another example is that in pharmacogenetic studies (studies that investigate how our genes could influence the response to a medication) of warfarin, the ability to accurately predict a suitable warfarin dose as soon as possible is a priority for the NHS and its patients; in particular, to prevent haemorrhagic (bleeding) events; the most feared and serious adverse event associated with warfarin treatment. Warfarin is a common, very effective anticoagulant (prevent the clotting of blood) used by up to 6% of the UK population at any one time. Some patients need as little as 0.5mg/day to maintain adequate anticoagulation, whilst others require more than 10mg/day. Because of this high variability, the International Normalised Ratio (INR); a measure of the blood's clotting capability; needs to be monitored regularly in patients being initiated onto warfarin, meaning a significant NHS cost burden and patient inconvenience. Although, it is identified that genetic factors could explain up to 60% of dose variability, the relationship between patient's genes and their response to warfarin such as the timing of haemorrhagic events and INR has not yet been fully understood due to the influence of the inherent relationship between the amount of warfarin prescribed (dose) and compliance with the prescription (adherence); eg. adherence worsening with increased dose due to haemorrhagic events.

To analyse such multidimensional data that commonly arise in clinical and public health research creates a pressing need to develop novel statistical models that describe the complex relationship between the biomarkers in relation to the timing of clinical events. The models need to take into account errors in the measurement of the biomarkers, irregularities and imperfectness of the measurements, and the variability observed between patients and over time, so as to be able to detect genuine patterns in the type of data observed in routine (observational) clinical practice. We propose to develop novel, but sufficiently flexible models that can implement in real time, in order to exploit biomarker data in providing accurate diagnostic and prognostic forecasts of clinical events enabling identification of individuals in whom early intervention may be warranted to prevent adverse clinical events. They will help the patient and their physician to gain a better understanding of the disease dynamics and ultimately take the most optimal decision of patient's (personalised) treatment at the current examination time point.

Due to current trends in medical practice towards personalised medicine, models that utilise all available information at a specific time point during patient's follow-up are proven to be quite valuable for physicians to gain a better understanding of the disease dynamics and ultimately take the most optimal decision of patient's treatment at that follow-up time point. We propose to develop a novel methodology to enable construction of prediction indices at the individual patient level combining the information at different time points on the multiple dimensions of longitudinal biomarkers and clinical endpoints.

Although limited number of studies were extended for multivariate dimensions of longitudinal biomarkers and event times focusing on specific aspects of motivating clinical studies, the focus was on estimating the association between the two processes, rather than producing individual and time-specific predictions of the event conditional on biomarker in combination. Most extensions rely on strong parametric assumptions regarding the two processes or are based on computationally complex techniques. A major drawback of above studies is the lack of user-friendly, readily available software that implement the models in real time. Further, the potential applications of this methodology within pharmacogenetic association studies, where true genetic effect on a patient's response to drugs is evaluated, have never been exploited.

Our proposal is built around above limitations within the development and application of multivariate joint modelling; aiming to develop a novel, but sufficiently flexible model by exploiting the underlying dependencies between multivariate dimensions of longitudinal and event time outcomes through correlated random effects; discuss a novel application of the proposed model in pharmacogenetics studies, and implement software in R and develop a training package to ensure practical relevance of models in mainstream medical research.

Our project will lead to novel methodologies that allow researchers to identify patients at higher risk of developing a particular disease, condition or adverse effect sooner than is currently achieved, so will impact on patients, clinicians and for health care costs as well as on statisticians.

Short-term impact (1-3 years and beyond)
Clinicians and Patients: The proposed models will be used to provide accurate diagnostic and prognostic forecasts of clinical events that enable identification of individuals in whom early intervention may be warranted to prevent adverse clinical events. They will help the patient and their physician to gain a better understanding of the disease dynamics and ultimately take the most optimal decision of patient's care at the current examination time point. In close collaboration with two leading clinicians we apply the proposed models in two priority areas of clinical management of the NHS patients. Our recommendations will strengthen the evidence-based health decisions; In sepsis, we inform clinical researchers on the association between biomarkers and the onset of sepsis, whether the biomarkers in combination improves the diagnosis and a time-specific optimal prediction rule for the diagnosis; In warfarin treatment, we provide more accurate details on the influence of warfarin exposure on association between genetic predictors and clinical outcomes of therapeutic INR and adverse events of warfarin.

Statisticians: Despite the development of numerous methodologies for complex data, simple statistical approaches remain widespread among applied statisticians, leading to inefficient analysis and the potential for important predictive relationships to be missed. One reason for this is that software to implement methods in real time is not generally available. Our work on software and dissemination will have direct impact on the way researches analyse such joint data structures. We update our current software package 'joineR' by adding a number of new functions to implement the proposed work. The software will be freely available for any operating system.

Researchers in statistical genetics: The novel application of the methodology within pharmacogenetic association studies will have an immediate impact on the way researchers in statistical genetics combine the information on patient's response to a treatment and their genetic information to provide a tailored and personalised approach to the patient's treatment.

Medium and long-term impact (2-4 years and beyond)
Statisticians and clinical researchers: The proposed work will also develop a platform for applied statisticians to bring in new clinical applications of the methodology. We update our current website for the joint modelling with information about our latest research activities to facilitate discussions and disseminate research outputs to a broader audience. The statistical methodologies, applications and clinical findings that will be disseminated via MRC NWHTMR, various meetings (eg. RSS medical section), workshops, conferences and publications will have long-term impact on the way data being analysed.

NHS policies and costs: The proposed methodologies will make it possible to understand the disease dynamics more accurately and to take the most optimal decision of patient's treatment at the current examination time point. This will allow reducing the cost burden to the NHS; In terms of patients being initiated onto warfarin, who need to be monitored regularly to prevent severe haemorrhage due to limited understanding of the disease dynamics, meaning a significant NHS cost burden and patient inconvenience; The challenges associated with sepsis in the UK are particularly concentrated on intensive care units with over 20,000 cases annually in England and Wales alone. Early detection and timely therapeutic intervention will ease the critical care burden to the NHS, and allow an improved outcome for patients with sepsis.