In this paper we demonstrate that routinely available intensive care unit (ICU) admission data is sufficient to build accurate machine learning workflows (AUC 0.86-0.89) for classifying a patient’s otherwise unknown ‘baseline’ for two blood results: hemoglobin and creatinine. The motivation for this work emerged from the frequent clinical encounter of acutely unwell patients with derangements in their laboratory blood results; specifically, hemoglobin and creatinine, both commonly deranged and clinically among the most instructive. Routinely, the first task to contextualise any derangement is to source historical blood results that allude to an acute or chronic trend. This may be a straightforward task for patients known to the admitting ICU’s affiliated healthcare system — on the presumption of an integrated electronic health record (EHR). However, faced with the reality of fragmented healthcare systems, quick reference to prior blood tests remains challenging. Separately, many previously 'well' patients may present needing intensive care support as part of a first-ever encounter with healthcare services, and therefore have no prior blood work to refer to. 

In the absence of centralised systems that show trends of blood results across long term follow up, we instead use clinical intuition in day-to-day practice. This form of approximation is used on a daily basis to guide our decision-making; if a patient admitted with a gastrointestinal bleed is stable from a haemodynamic standpoint, a low haemoglobin is interpreted differently, as may a patient’s elevated creatinine in the context of other markers of chronic renal disease. These assumptions of 'baseline' or close-to-baseline may lack objectivity -- and therefore consistency -- across different clinicians, ultimately causing significant variability in patient care. This is where algorithms such as the one described in our paper could add value. Clinical judgment and experience remains central to decision making for acutely unwell patients, but our work shows that there exists the opportunity to further enhance decision making through data-driven objectivity. Faced with a decision to transfuse or acutely consider renal replacement for a patient, the clinical community is heavily reliant on corresponding blood results for guidance. However, in the presence of doubt, deciding on how close a patient might be to their baseline levels could be unintentionally biased. Integrating the ability to highlight the likeliest baseline value within EHRs could lend an unrealised asset in supporting these decisions.

This work adds to the ever mounting body of evidence that machine learning applied to routine clinical data can unearth previously unimaginable, clinically useful insights. Once machine learning workflows such as the one outlined are trained and validated across further datasets of different populations, the promise of such formulaic decision aids could realize better consistency of decisions, many of which can carry significant risk. In the same way we already use an estimated glomerular filtration rate, which draws on four routinely available variables to generate an output, the same approach can be applied to further leverage data we already have readily available to predict baseline blood results across a diverse panel of tests.