Robotic process automation to identify patients at high risk for perioperative myocardial infarction or injury: a prospective, blinded, paired reader-controlled single-centre study

Abstract

Background: Although current guidelines recommend active surveillance for perioperative myocardial infarction, injury, or both in high-risk patients, implementation remains limited in most institutions worldwide because of a lack of resources.

Methods: We hypothesised that robotic process automation (RPA), a software technology that enables virtual bots to replicate human tasks within digital systems, could accurately replace experienced clinical staff. Manual screening by experienced clinical staff and RPA screening were carried out simultaneously and blinded to identify high-risk patients eligible for active surveillance for myocardial infarction/injury, according to predefined screening criteria. Discrepant identification was reviewed by an independent clinician blinded to the origin of the identification, generating a reference standard classification of paired reader-controlled patients to investigate the primary diagnostic endpoint: relative true positive fraction.

Results: In 660 participants (median age 60 yr, interquartile range 42-73 yr, 54.8% female), 77/660 (12%) were eligible for active surveillance for perioperative myocardial infarction or injury according to the reference standard classification. RPA screening achieved 75 (97%) true positive identifications, compared with 63 (82%) identified from manual screening (relative true positive fraction: 1.19, 95% confidence interval 1.08-1.32, P=0.004). The number needed to screen to identify one additional true positive using RPA screening was 6. RPA screening had a sensitivity of 0.97 (0.91-0.99), compared with 0.82 (0.72-0.89) for. Both approaches had high specificity (RPA screening: 0.98 [0.97-0.99], compared with manual screening: 1.0 [0.99-1.00]). The estimated annual cost of RPA screening was 81% lower compared with manual screening.

Conclusions: RPA screening was superior to standard-of-care manual screening by experienced clinical staff in identifying patients at high risk for perioperative myocardial infarction or injury.

Clinical trial registration: NCT02573532.

Keywords: active surveillance; myocardial injury after noncardiac surgery; perioperative care; perioperative myocardial infarction; perioperative myocardial injury; robotic process automation.

Fig 1

Screening criteria of patients scheduled for surgery. (a) High-risk and inclusion criteria for PMI surveillance. All criteria must be met. (b) Low-risk and exclusion criteria. Meeting any one of these is sufficient to exclude a participant from PMI surveillance, even if inclusion criteria are fulfilled. PMI, perioperative myocardial infarction or injury.

Fig 2

Reference standard classification. (a) Flowchart of participant classification and adjudication process. In cases where RPA screening and MS agreed, participants were directly included or excluded based on that consensus. Discrepant identifications were reviewed in a paired reader control process by an independent clinician blinded to the screening method. Finally, the reference standard classification was based on agreement among at least two of the three sources: RPA screening, MS, and the independent clinician. (b) Alluvial plot illustrating agreement and discrepancies between all methods. MS, manual screening; Ref, reference standard; RPA, robotic process automation.

Fig 3

Diagnostic accuracy of RPA screening vs manual screening for identifying participants at high risk for PMI. Comparative forest plot of sensitivity and true positive identification rates between RPA screening and MS, displayed as relative true positive fraction (i.e. the number of true positive identifications by manual screening and RPA screening divided by the number of true positive cases according to the reference standard classification). All estimates are presented with corresponding 95% confidence intervals. The striped line indicates the predefined non-inferiority margin of 0.9. MS, manual screening; PMI, perioperative myocardial infarction/injury; RPA, robotic process automation.

Fig 4

Keyword-based participant identification by robotic process automation screening. Bar plot of high-risk and eligible participants identified by robotic process automation screening based on keyword extraction from preoperative anaesthesia assessments and medical records. PMI, perioperative myocardial infarction/injury.