It is difficult to provide evidence for how accurately risk can be predicted because there is a lack of empirical data connecting patient cues to outcomes (e.g. Stein 2002; DOH, 2007; RCP, 2008) and because what data does exist tend to focus on cues considered in isolation rather than on their patterns of occurrence and interactions. Hence even tools with a better grounding in actuarial evidence have been shown to over-predict rare occurrences and have only moderate accuracy (Snowden et al 2007; Coid et al 2009). Studies of suicide prediction specifically have shown high levels of false positives (Bouch and Marshall, 2005; Harriss and Hawton, 2005). This has caused doubt about the efficacy of actuarial tools when used on individual cases (Godin, 2004; Holdsworth & Dodgson, 2003). It was also a reason why the UK Department of Health (DH 2009) and the Royal College of Psychiatrists (RCP 2008) advocated combining actuarial approaches with structured clinical judgement: the approach taken by GRiST. However new research (e.g. Bakst et al. 2010; Jiang et al. 2010) and evidence synthesis (e.g. Bowers et al. 2010) is beginning to provide more information about cue clusters associated with high risk of suicide. Analysis of the GRiST database will contribute to the evidence about this and about other risks too
Regarding the use of GRiST for predicting risks, its current objective is more about supporting clinical risk judgements and the associated advice rather than trying to output precise probabilities or some associated overall risk score. Scores and probabilities are extremely hard to produce with any accuracy or confidence, partly due to the current paucity of data but also because suicide, for example, is a very rare event. Addressing predictive validity of tools prospectively (i.e. testing how accurate a risk prediction is after it has been made) is difficult because high risks will instigate associated interventions to reduce them and any connection between risk assessment and risk prediction is disrupted, which is, of course, the aim of good clinical practice.
At present, no risk tools have high accuracy but they still have an important role because they identify relevant risk information that is known to increase the risks, albeit not by any precisely known amount, and thus can suggest interventions that will certainly have impact on preventing tragic outcomes. In short, risk-related data should be collected because they guide decisions about the most appropriate care a person should receive, irrespective of whether or not the data accurately predict potential risk behaviours. Furthermore, this information will, in time, begin to fill in the gaps in knowledge about how patient cues link to risk probabilities. Unfortunately, risk data are currently not collected very systematically, are often in formats that are hard to analyse, and may not be easily accessed by computers. GRiST was designed to tackle all three of these problems by having a comprehensive data set, recording the data in highly structured formats that explicitly encode the relationships between risk data and concepts, and by linking the input data to structured clinical judgements that are automatically recorded in an electronic database for analysis.