Towards Clinically Useful Expert Systems in Critical Care: Locally Managed Interpretation of Arterial Blood Gas Data

Abstract

Clinical pathology tests generate large volumes of data for critical care clinicians, and ”data overload”, can impede clinical decision making. Bedside arterial blood gas (ABG) analysis is compounding this problem. Expert systems (ES) have great potential to support clinicians' work, but few ES are in routine use. The complexity of knowledge acquisition (KA) leads to poor productivity in the development phase and requires specialist knowledge engineering (KE) skills. This has been termed the knowledge acquisition “bottleneck”. We have devised a novel algorithm, ripple-down rules (RDR), that essentially eliminates the KE bottleneck. The Pathology Expert Interpretative Reporting System (PEIRS), an RDR ES, interprets a broad range of pathology reports (including thyroid function tests, ABG, cardiac enzymes and a range of hormone assays) and has been in routine use for over 4 years. With over 2100 rules, PEIRS is one of the few large medical ES in routine use, and apparently the only large ES built entirely by a pathologist without specialist KE skills or support. A key finding from PEIRS was that its simple KA strategy enabled localisation of its knowledge, e.g. local protocols for interpretation and management. This localisation was essential for gaining the acceptance of the system by the pathologists.

While PEIRS successfully automated the interpretation of ABG reports, extension to bedside decision support was limited by its restriction of a single interpretative comment per report. A new algorithm, multiple-classification RDR (MC-RDR), addresses this issue. We compared conventional RDR and MC-RDR by building prototype ES for ABG interpretation. Data included pH, PCO₂, PO₂, calculated bicarbonate, base excess and patient's age on up to five sequential specimens. Interpretations such as “resolving respiratory acidosis, metabolic compensation” were based on local criteria, e.g. “severe hypoxaemia” if PO₂ is less than 59 mmHg. After 262 cases, MC-RDR was far more compact than RDR (262 vs. 442 total rule conditions), simpler to maintain (109 vs. 184 rules; 22 vs. 129 comments) and matured earlier (84% vs. 30% accuracy for the last 50 cases). The KA task for the local expert was marginally more complex with MC-RDR, but this was more than offset by the substantial overall reduction in KA requirements.

We have shown that MC-RDR is more compact than RDR, and KA is far more efficient. Its rapid maturation supports application in routine clinical care and would suit an embedded decision support system for an ABG analyser. By enabling localisation, we believe that MC-RDR redefines the role of medical ES and that real-time MC-RDR ES will soon be helping critical care clinicians reduce data overload in their daily practice.