Skip to main content

Table 2 Participant-level diagnostic label precision and recall estimates against reference standard (i.e. expert physician adjudications based on the complete EMR)

From: Developing automated methods for disease subtyping in UK Biobank: an exemplar study on stroke

Stroke subtype Precision (i.e. positive predictive value)
(95% CI)
Recall (i.e. sensitivity)
(95% CI)
From codes (based on previous work [2]) From automated method From codes (based on previous work [2]) From automated method
ICH 42% (31–54%)
(11/26)
89% (52–100%)
(8/9)
100% (72–100%)
(11/11)
89% (52–100%)
(8/9)
SAH 71% (54–83%)
(17/24)
82% (57–96%)
(14/17)
100% (80–100%)
(17/17)
82% (57–96%)
(14/17)
IS 83% (75–89%)
(73/88)
73% (65–81%)
(91/124)
49% (41–57%)
(73/149)
64% (56–72%)
(91/142)
IS (including cases with an unspecified subtype assigned as IS) 80% (76–83%)
(147/184)
77% (71–83%)
(141/182)
99% (95–100%)
(147/149)
99% (96–100%)
(141/142)
  1. ICH, intracerebral hemorrhage; SAH, subarachnoid hemorrhage; IS, ischemic stroke; IS (including cases with an unspecified subtype assigned as IS) = all cases where a stroke subtype could not be assigned with automated methods or where the code was unspecified for a stroke subtype were assumed to be ischemic stroke; Precision = positive predictive value (proportion of true-positive cases among all cases). Recall = sensitivity (proportion of all true-positive cases in the population identified). Absolute numbers of cases provided in brackets. The dataset used for the precision and recall calculation from codes in our previous work [2] included a total of 225 participants with a stroke code. The dataset used for the precision and recall calculation from automated method in this study includes a total of 207 participants with a stroke code. The 207 are a subset of the 225 participants with a stroke code who also had a relevant clinical brain scan report available. 18 participants among the 225 participants did not have a brain scan available and were hence excluded from this study