Page 45 - Read Online

Shapey et al. Art Int Surg 2023;3:1-13 https://dx.doi.org/10.20517/ais.2022.31 Page 3

Perioperative prediction
Using perioperative data to pre-empt postoperative complications is not a new concept, and is fundamental
to contemporary management of postoperative surgical patients. At an elementary level, clinicians use
mental models such as recognition primed decision making, critical decision methods, and data frame
theory [11-13] . These models of decision making are the framework for what is more commonly described as
“expertise” or “experience”. Such mental models, although often correct, are open to error, misuse or
misdirection [11-13] . In the search for additional data in support of a specified hypothesis (sensemaking),
individuals may be drawn along an erroneous path and misattribute data to the wrong association or cause.
It is easy to fall into the cognitive trap of “explaining away” the association between poor outcomes and
technical errors, or to over-interpret the significance of an adverse event in a patient whose morbidity may
have little to do with the surgeon themselves. The potential value of ML, therefore, to objectively identify
anomalous data and high-risk physiological patterns is of great importance. Cognitive bias may also lead
surgeons to change a technical approach when no change is warranted, and vice versa.

One method of pre-emptively identifying and pro-actively addressing potential complications is the use of
[14]
electronic app-based clinical algorithms, as reported by the PORSCH trial in pancreatic surgery . In this
randomised controlled trial of best practice after pancreatic resection in the Netherlands, algorithm-based
care was used to determine when to perform an abdominal CT, radiological drainage, start antibiotic
treatment, and remove abdominal drains. The algorithms described in this study represent at a human level
what computers seek to achieve at a digital level. The value of algorithms of optimal perioperative care is
illustrated by a significantly lower rate of the primary outcome (bleeding that required invasive
intervention, new-onset organ failure, and death either during admission or within 90 days after resection)
in the intervention group utilising the algorithm (adjusted RR 0.48, 95%CI: 0.38-0.61; P < 0.0001). It is also
important to consider how ML algorithms could improve the prediction of postoperative complications
above and beyond existing optimal systems and human-derived algorithms. Moreover, defining the key
outcome of interest, e.g., failure to rescue rather than new-onset organ failure per se, is of paramount
importance in shaping the way that ML will interact with clinical practice.

Modified Early Warning Scoring (MEWS) systems exist to identify and pre-empt clinical deterioration, and
are based on basic physiological parameters such as heart and respiratory rate, blood pressure, oxygen
saturation and requirement, and neurological status [15-16] . In many healthcare systems, MEWS systems can
be set at certain thresholds to trigger pre-determined actions by clinical staff, for example, the automated
review of a patient by a critical care outreach team. Such systems have been shown to have a beneficial
impact on medical patient care by reducing the rate of in-hospital cardiac arrest [17-19] . The absence of
individual patient context to the interpretation of MEWS data outputs (e.g., heart rate and beta-blockade or
athleticism) represents a critical limitation, as does the non-identification of critical junctures that arise
from reviewing isolated data outputs rather than appreciating the subtleties of data trends (e.g., swinging
pyrexia). ML could help address the deficiencies in existing systems: (a) by identifying anomalous data that
does not trigger an automated or human system; (b) by relating biomarker data to electronic health and
prescribing records; and (c) by alerting clinicians to concerning clinical note entries through free-text
associations.

Currently, the practical application of ML to perioperative care is limited by multiple stumbling blocks.
These include: (a) the accuracy of alerts and the potential of spurious data to divert attention; (b) real-time
delivery of alerts in a manner that could change clinical practice; and (c) convergence of data points and
gate-keeping over which data ought to be considered relevant. In due course, these limitations could each be
addressed by the regular auditing and quality control of ML systems, by automating real-time calculations
and subsequent alerts to accompany each new piece of data, and by utilising multi-faceted and integrated
electronic patient records.