Table 2.

Positive predictive values of the natural language processing algorithms.

	Pre-pandemic positive predictive value (95%CI, No. annotated records)a	Post-pandemic positive predictive value (95%CI, No. annotated records)a
Main algorithm used to detect hospitalisations caused by suicide attempts (hybrid machine learning and rule-based approach)	0.85 (0.76–0.91, 85)	0.86 (0.76–0.92, 77)
Alternative algorithm used to detect hospitalisations caused by suicide attempts (rule-based approach)	0.51b	0.52b
Algorithm used to detect social isolation risk factor	1.00c (0.89–1.00, 30)	0.96c (0.79–0.99, 23)
Algorithm used to detect domestic violence risk factor	0.96c (0.79–0.99, 23)	0.96c (0.80–0.99, 24)
Algorithm used to detect sexual violence risk factor	0.84c (0.70–0.93, 38)	0.96c (0.82–0.99, 28)
Algorithm used to detect physical violence risk factor	0.86c (0.69–0.95, 29)	0.96c (0.82–0.99, 28)
Algorithm used to detect suicide attempt history risk factor	0.93c (0.78–0.98, 29)	0.83c (0.66–0.93, 30)

^aWilson score intervals were used to compute the 95%CIs.

^bFor the estimation of the performances of the alternative rule-based SA-classification algorithm, additional stays were drawn randomly among the stays that were classified as SA-caused by the rule-based algorithm but not by the hybrid algorithm—see Supplement. The Wilson score intervals could not be computed using this method.

^cFor the estimation of the performances of the risk factors algorithms, stays were drawn randomly among those that were labelled positive both by the main SA-classification algorithm and by the risk factors-classification algorithms.