Predicted risk of harm versus treatment benefit in large randomised controlled trials

Abstract

Most drugs come with unwanted, and perhaps harmful, side-effects. Depending on the size of the treatment benefit such harms may be tolerable. In acute stroke, treatment with aspirin and treatment with alteplase have both proven to be effective in reducing the odds of death or dependency in follow-up. However, in both cases, treated patients are subject to a greater risk of haemorrhage – a serious side-effect which could result in early death or greater dependency. Current treatment licenses are restricted so as to avoid treating those with certain traits or risk factors associated with bleeding. It is plausible however that a weighted combination of all these factors would achieve better discrimination than an informal assessment of each individual risk factor. This has the potential to help target treatment to those most likely to benefit and avoid treating those at greater risk from harm. This thesis will therefore: (i) explore how predictions of harm and benefit are currently made; (ii) seek to make improvements by adopting more rigorous methodological approaches in model development; and (iii) investigate how the predicted risk of harm and treatment benefit could be used to strike an optimal balance. Statistical prediction is not an exact science. Before clinical utility can be established it is essential that the performance of any prediction method be assessed at the point of application. A prediction method must attain certain desirable properties to be of any use, namely: good discrimination – which quantifies how well the prediction method can separate events from non-events; and good calibration – which measures how close the obtained predicted risks match the observed. A comparison of informal predictions made by clinicians and formal predictions made by clinical prediction models is presented using a prospective observational study of stroke patients seen at a single centre hospital in Edinburgh. These results suggest that both prediction methods achieve similar discrimination. A stratified framework based on predicted risks obtained from clinical prediction models is considered using data from large randomised trials. First, with three of the largest aspirin trials it is shown that there is no evidence to suggest that the benefit of aspirin on reducing six month death or dependency varies with the predicted risk of benefit or with the predicted risk of harm. Second, using data from the third International Stroke Trial (IST3) a similar question is posed of the effect of alteplase and the predicted risk of symptomatic intracranial haemorrhage. It was found that this relationship corresponded strongly with the relationship associated with stratifying patients according to their predicted risk of death or dependency in the absence of treatment: those at the highest predicted risk from either event stand to experience the largest absolute benefit from alteplase with no indication of harm amongst those at lower predicted risk. It is concluded that prediction models for harmful side-effects based on simple clinical variables measured at baseline in randomised trials appear to offer little use in targeting treatments. Better separation between harmful events like bleeding and overall poor outcomes is required. This may be possible through the identification of novel (bio)markers unique to haemorrhage post treatment.