Machine learning in prognosis research

Machine learning offers an alternative to the methods for prognosis research in large and complex datasets and for delivering dynamic models of prognosis. Machine learning foregrounds the capacity to learn from large and complex data about the pathways, predictors, and trajectories of health outcomes in individuals. This reflects wider societal drives for data-driven modelling embedded and automated within powerful computers to analyse large amounts of data. Machine learning derives algorithms that can learn from data and can allow the data full freedom, for example, to follow a pragmatic approach in developing a prognostic model. Rather than choosing factors for model development in advance, machine learning allows the data to reveal which features are important for which predictions. This chapter introduces key machine learning concepts relevant to each of the four prognosis research types, explains where it may enhance prognosis research, and highlights challenges.

Electronic healthcare records and prognosis research

Electronic healthcare record (EHR) data, collected during the daily business of patient consultations and treatments, offer huge opportunities to expand the range and scale of prognosis research, in particular because of the real-time and continuous recording of potential prognostic factors and health-related events, and the amount of data and individuals involved. However, with these opportunities come challenges related to the size and complexity of EHR data. This chapter provides an overview of these issues.

Prognosis research in people with coronary heart disease

Coronary heart disease (CHD) outcomes have improved in recent decades because of better treatment, improved investigations, and better secondary prevention. The results of prognosis research have contributed to the development and evaluation of these new components of healthcare for CHD, but have also critically questioned traditional classifications of CHD, emphasized the importance of long-term outcomes in judging the success of healthcare in CHD patients, and highlighted the potential of risk stratification to guide better treatment decisions for individual patients with CHD. This chapter uses example studies to illustrate this story.

Predictors of treatment effect

A predictor of treatment effect is any factor or combination of factors (such as a patient characteristic, symptom, sign, test, or biomarker result) associated with the effect (benefit or harm) of a specific treatment in persons with a particular disease or health condition. Various terms are used across disciplines to refer to prediction of treatment effect, including treatment-predictor (treatment-covariate) interaction, effect modification, predictive (as opposed to prognostic) factors (in oncology), or moderation analysis. This chapter reviews principles of the design of studies of treatment effect predictors, such as exploration of treatment-predictor interactions in randomized trials and the importance of replication of such estimates using data from multiple trials. The application of predictors of treatment effect in practice for matching individuals or subgroups to specific treatments is introduced as one type of stratified care, and the need for impact studies to investigate whether stratified care leads to better outcomes and improved efficiency of healthcare is highlighted.

A framework for prognosis research

This chapter introduces the PROGRESS framework, which describes four types of prognosis research, each addressing different questions. The four types concern: studies of overall prognosis (the average outcome, or outcome risk, in people with a particular health condition, in the context of the nature and quality of current care); prognostic factors (characteristics associated with changes in the average outcome, or outcome risk, across individuals); prognostic models (development, validation, and impact evaluation of statistical models, incorporating multiple prognostic factors for use in clinical practice to predict an individual’s outcome value or to estimate their outcome risk); and predictors of treatment effect (characteristics that predict whether an individual responds to a particular treatment or not). Examples of each type are given to illustrate the framework.

Individual participant data meta-analysis of prognosis studies

An alternative approach to meta-analysis of aggregate data from published prognosis research (as addressed in Chapter 9), with its challenges of heterogeneity and lack of information, is to conduct meta-analysis of individual participant data (IPD), that is, the original raw data of the individuals who are included in the primary prognosis studies. The approach is increasingly feasible as data sharing and open-access data become more popular, and the chapter highlights why they offer enormous advantages for a robust and meaningful evidence synthesis of prognosis studies. In particular, better prognostic models can be developed and directly validated across multiple settings, and power is increased to detect genuine predictors of treatment response. Key steps in such an IPD meta-analysis are explained, including practical guidance on how to obtain, handle, and synthesize data, and what potential challenges may be encountered.

Systematic reviews and meta-analysis of prognosis research studies

Systematic reviews and meta-analyses identify, evaluate, and summarize prognosis research studies and their findings. The chapter provides a guide to the key components and methods for conducting a systematic review and meta-analysis for each of the four types of prognosis studies. The CHARMS checklist is introduced as a guide to identifying clear review objectives and design, and to extracting the relevant information from each included study. Many existing prognosis studies are at high risk of bias, because (for example) of selective recruitment and reporting. Tools for examining quality of studies are discussed—the QUIPS for prognostic factor research and PROBAST for prognostic model research. The statistical principles of meta-analysis are described, and the key statistics that can be synthesized are outlined. Challenges are identified, such as the potential for publication bias and substantial heterogeneity in published prognostic factor cut points and methods of prognostic factor measurement. Despite these challenges the chapter emphasizes the crucial importance of prognosis reviews for evidence-based guidelines and clinical decision making.

Overall prognosis research

Overall prognosis research concerns the description of average future outcomes of groups of people with a certain disease or health condition in the context, time, and setting of current healthcare. This chapter describes how overall prognosis is estimated among people with a defined health condition in relation to relevant health outcomes. Study design, from newly designed prospective cohorts to cohorts embedded in routine healthcare data, is discussed. The value of information derived from overall prognosis research for patients and for healthcare professionals, policymakers, and funders, is considered, particularly in relation to decision making in healthcare practice and to monitoring healthcare outcomes for policymaking. Wider roles of overall prognosis estimation in informing other types of prognosis research, the design and interpretation of treatment effectiveness studies, understanding the consequences of using new diagnostic tests, and identifying unintended benefits or harms of treatment, are described.

Prognosis research in people with traumatic bleeding

Prognosis research has played a major role in the development of approaches to the management of trauma. This is because of the need to identify those people who have a poor immediate prognosis if untreated and because of the many settings where choices have to be made on which patients to focus life-saving resources. This need for evidence-based triage based on prognostic information is particularly true for the problem of traumatic bleeding, and this chapter details the development and validation of a prognostic model and predictors of benefits or harms of treatment.

Prognosis research in people with low back pain

Prognosis has dominated recent low back pain (LBP) research because of the lack of disease pathological explanations of LBP that lead to safe and effective treatments in many patients; the hazards of overdiagnosis and overtreatment; and the potential for beneficial outcomes in patients if treatment approaches are carefully matched to the likelihood of recovery, recurrence, or persistence, or the likely effect of specific treatments. This chapter uses examples from each of the four types of prognosis research to illustrate how prognosis research has contributed to understanding LBP and provided evidence to inform classification and treatment of patients with LBP.