Analyzing historical and future acute neurosurgical demand using an AI-enabled predictive dashboard

Abstract

Characterizing acute service demand is critical for neurosurgery and other emergency-dominant specialties in order to dynamically distribute resources and ensure timely access to treatment. This is especially important in the post-Covid 19 pandemic period, when healthcare centers are grappling with a record backlog of pending surgical procedures and rising acute referral numbers. Healthcare dashboards are well-placed to analyze this data, making key information about service and clinical outcomes available to staff in an easy-to-understand format. However, they typically provide insights based on inference rather than prediction, limiting their operational utility. We retrospectively analyzed and prospectively forecasted acute neurosurgical referrals, based on 10,033 referrals made to a large volume tertiary neurosciences center in London, U.K., from the start of the Covid-19 pandemic lockdown period until October 2021 through the use of a novel AI-enabled predictive dashboard. As anticipated, weekly referral volumes significantly increased during this period, largely owing to an increase in spinal referrals (p < 0.05). Applying validated time-series forecasting methods, we found that referrals were projected to increase beyond this time-point, with Prophet demonstrating the best test and computational performance. Using a mixed-methods approach, we determined that a dashboard approach was usable, feasible, and acceptable among key stakeholders.

Figure 1

Data acquisition, processing, analysis and visualization pipeline (CNN-LSTM convolutional neural network—long short-term memory, STL seasonal and trend decomposition using Loess, ARIMA automated regression integrated moving average).

Figure 2

Age and sex distribution of referrals presented by diagnostic classification (SDH subdural hemorrhage, M male, F female).

Figure 3

Referral classification. Referral proportion by diagnostic classification (A), urgency (B), referrer type and seniority (C) (SDH subdural hemorrhage, F1/F2 intern/foundation year ½, SHO senior house officer, SpR specialist registrar/resident, ANP advanced nurse practitioner, CNS clinical nurse specialist, GP general practitioner).

Figure 4

Referral triage decisions. A Sankey plot illustrating decisions made by the on-call registrar/resident sorted by diagnostic classification (blue bars). Triage decisions are grouped according to whether the patient was accepted (green bars), further information was requested (purple bars), advice or conservative management was suggested (orange bars), an additional neurosurgical review was needed to make a decision (yellow bars) or whether the referral was rejected (red bars) (MDT multidisciplinary team meeting).

Figure 5

Referral heatmap. Referral volumes sorted by day and time for all (A) and for the four highest referring diagnostic categories (B).

Figure 6

Geographic referral visualization. (A) Map of referring sites (red dots) to the neurosurgical center from across the U.K. between March, 2020 and October 2021. (B) Northern Greater London referral catchment area with referring sites (red circles) size proportional to referral volume. The five highest volume main referring sites are highlighted with black borders. ^†Denotes the approximate location of the receiving neurosurgical center.

Figure 7

Time-series algorithm performance and prediction of future referral volume. (A) Evaluation of time-series forecasting algorithms using fivefold block cross-validation (CV) and train-test split (Test) using mean absolute error (MAE), mean absolute predictive error (MPE) and root mean squared error (RMSE) with 1, 4 and 12 week forecasting periods. Algorithm legend: computational time (t) taken to run onefold of each algorithm (STL seasonal and trend decomposition using Loess, CNN-LSTM convolutional neural network—long short-term memory, ARIMA automated regression integrated moving average). (B) Timeline of weekly referral volumes plotted since the start of the Covid-19 pandemic with in-sample forecasting (prediction: yellow dashed line) and 12-week out-of-sample forecasting (future: red dashed line) determined by Prophet with 95% confidence intervals shown in gray.