Fred Vivian: Increasing the use of data to guide decision-making

As clinicians we assimilate data to make decisions regarding patient care. Through rehearsed pattern recognition we incorporate large quantities of data to diagnose and formulate management plans. Since the scientific revolution, we have used statistical models to analyse clinical data in medical research.

However, we use very little statistical modelling or computerised data analysis in our clinical practice. Instead we often rely on a single clinician sitting in a room, thinking deeply. This system is highly vulnerable to variations in clinician experience, skill or the time of day (studies have shown clinicians order fewer tests, prescribe fewer antibiotics and are less likely to offer surgery later in the day) [1-3].

At present large language models such as chat-GPT or those trained more specifically on medical data-sets such as Google’s Med-PaLM can do fairly well in medical exams, but (at least for now) are too unreliable for real world conditions [4]. So, how can we safely improve care now, without relying on our brains alone to crunch data.

We have long used risk scores to support clinical judgement such as a Wells score or NEWS score. Data-led prioritisation (DLP) uses clinical data to risk stratify and then prioritise patients on the waiting list.

Factor 50, whose DLP product was developed in conjunction with Guy’s and St Thomas’ NHS Trust, initially looked at prioritising diabetes patients analysing data including HbA1c and A&E attendances to risk stratify patients. 

The initial cohort identified hidden risk that was pro-actively addressed and increased the use of PIFU [5]. The study also demonstrated that non-Caucasian and more socio-economically deprived patients were more likely to be in the high risk category. So this offers a route to reducing health inequalities. Another benefit of DLP is its ability to identify inter-clinician variability in risk management. This can then be analysed and, where this is unwarranted, be used as an educational tool. 

While at present these risk scores are fairly crude, in the near future, machine learning will be able to refine these to improve their accuracy to the point that they out-perform clinicians in judging clinical risk. With increased use of remote monitoring and patient reporting of deterioration, this can be further enhanced to identify patients in need of escalation for early review.

DLP is one tool that will enable transformation of outpatient care to a much more dynamic and reactive system, using clinicians’ time more smartly to see the patients who will benefit most.