Predicting ophthalmic clinic non-attendance using machine learning: Development and validation of models using nationwide data

Background

Ophthalmic clinic non-attendance in New Zealand is associated with poorer health outcomes, marked inequities and costs NZD$30 million per annum. Initiatives to improve attendance typically involve expensive and ineffective brute-force strategies. The aim was to develop machine learning models to accurately predict ophthalmic clinic non-attendance.

Methods

This multicentre, retrospective observational study developed and validated predictive models of clinic non-attendance. Attendance data for 3.1 million appointments from all New Zealand government-funded ophthalmology clinics from 2009 to 2018 were aggregated for analysis. Repeated ten-fold cross validation was used to train and optimise XGBoost and logistic regression models on several demographic and clinic-related variables. Models developed using the entire training set were compared with those restricted to regional subsets of the data.

Results

In the testing data set from 2019, there were 407 574 appointments (median [range] age, 66 [0–105] years; 210 365 [51.6%] female) with a non-attendance rate of 5.7% (n = 23 309 missed appointments), XGBoost models trained on each region's data achieved the highest mean AUROC of 0.764 (SD 0.058) and mean AUPRC of 0.157 (SD 0.072). XGBoost performed better than logistic regression (mean AUROC = 0.756, p = 0.002). Training individual XGBoost models for each region led to better performance than training a single model on the complete nationwide dataset (mean AUROC = 0.754, p = 0.04).

Conclusion

Machine learning algorithms can predict ophthalmic clinic non-attendance with relatively basic demographic and clinic data. These findings suggest further research examining implementation of such algorithms in scheduling systems or public health interventions may be useful.     

Intracranial tuberculoma. Evaluation and treatment

We used computerized tomography (CT) to screen 83 patients with partial seizures who had 1) increased intracranial pressure; 2) evidence of systemic tuberculosis; or 3) focal neurological deficit. We found intracranial tuberculoma in 20 (24%). In addition, of 55 patients with tuberculous meningitis, 12 had associated tuberculoma. The image morphology on CT scans showed ring lesions (14), discs (10), and irregular coalescing masses (8). Perilesional edema frequently was present. After appropriate antitubercular therapy, clinical outcome for seizures, neurological deficit, and increased intracranial pressure was satisfactory. CT lesions regressed within 12 weeks of inauguration of medical therapy in most patients. Medical management of tuberculoma is advocated, with surgery limited to those in whom such treatment is ineffective.     

Towards pain free cardiac surgery - high thoracic epidural analgesia

Epidural analgesia is a well-established analgesic technique, providing complete or near complete perioperative analgesia for major surgery. Epidural analgesia may also confer additional benefits during the perioperative period, including attenuation of the stress response, reduction in procoagulation tendency, and reduction in myocardial ischaemia and infective complications. These additional benefits are an advantage in patients with ischaemic heart disease. We discuss the use of high thoracic epidural anaesthesia in cardiac surgery, outlining potential benefits and risks. Our experience with epidural analgesia in over 300 patients undergoing cardiac surgery is briefly described.