Diagnostic errors are common in clinical practice. However, it is not known how to estimate the probability that a hospitalized patient has an incorrect, delayed or missed diagnosis. The purpose of our project was to create a framework that could help us understand how to estimate the chance of error at the time of admission and how to re-estimate that risk over time during hospitalization. In order to do this, we first extracted a list of risk factors that have been reported in the literature to be associated with diagnostic errors. We then used systems engineering methods to analyze cases of patients that we thought had a higher probability of diagnostic error. We compared what was reported in the literature with our structured analysis. We finally compiled a list of risk factors that could be used to design a computer program (machine learning algorithm) that once validated, could potentially calculate the probability of diagnostic error in real time.
To date, there are no recommendations on how to estimate the risk of diagnostic errors (DEs) in the inpatient setting. We created a framework to understand the risk of DE as a function of time. We conducted a literature review of risk factors for DEs. We then used root-cause and failure mode and effect analysis to identify potential risk factors for DEs. We evaluated cases from a random sample of general medicine patients with “e-triggers”; events thought to be associated with DEs (e.g. death during hospitalization, transfer to ICU within 72hrs of admission). Finally, we used an iterative approach to construct a consensus framework. We compared the identified risk factors from our retrospective analysis to those reported in the literature.
Our framework uses EHR data to assess baseline risk at the time of admission based upon demographic factors, high-risk comorbidities, patient complexity and evidence of clinical deterioration in the outpatient setting. This risk is modified dynamically by new clinical data and electronic events during hospitalization. We propose a compiled list of potential covariate candidates to be initially included in the training phase of a machine learning model that could potentially estimate diagnostic error risk in real time.