Background: Diagnostic errors (DEs) are common and can lead to preventable harm in hospitalized patients.[1] To address this problem as part of our AHRQ-funded Patient Safety Learning Laboratory, we characterized diagnostic process failures that contribute to DEs.[2,3] Next, we refined three interventions (Figure 1) that addressed common process failures and were embedded into our electronic health record (EHR)-integrated digital health platform for clinicians and patients.[4,5] The intervention components included a DE risk stratification algorithm,[6] a patient diagnostic questionnaire,[7] and 5-step diagnostic time-out (DTO).[8] In this study, we report preliminary estimates of DE rates before and after implementation of this preventative intervention using a stratified sample of cases from pre-defined cohorts of varying DE risk.
Methods: Due to the COVID-19 pandemic, we adjusted our implementation to include virtual DTO training sessions, brief animated videos, an educational podcast, and interactive “case of the month” newsletters. After excluding patients hospitalized less than 24 hours on general medicine or with a length of stay more than 21 days, we used random, stratified sampling to assemble four cohorts retrospectively at the end of the pre-implementation period (7/2019-3/2020), and prospectively during the post-implementation (1/2021 to 9/2021) period based on a priori trigger events. Cases in the high-risk cohorts included patients who were transferred to the ICU after 24 hours, died within 90-days of index hospitalization, or had other clinical trigger events (multiple consultants, new or worsening oxygen requirement, acute kidney injury, rapid response, etc.). Cases in the low-risk cohort did not have any of the above triggers. Each case was reviewed independently and reconciled by two hospitalist adjudicators trained in our structured DE case review process which used validated instruments (SAFER Dx) to assess DE outcomes.[9] All cases with discrepancies were reviewed by a tertiary panel of experts. All trained adjudicators (~20) also participated in the intervention. We compared outcomes using a weighted analysis and used propensity weighting to adjust effect size estimates for the entire cohort.
Results: Of 9147 available cases, 675 (339 pre, 336 post) were sampled in equivalent percentages from both periods: 100% of ICU transfers, 38.5% of deaths without ICU transfer, 7% of cases with other clinical triggers, 2.4% of cases with no triggers. Post-implementation, more patients (Table 1) were Hispanic, from lower socioeconomic status, or enrolled in Medicaid. The estimated rates (Table 2) of overall (harmful and non-harmful) DE and harmful DE attribute to the primary or a secondary diagnosis were 18.7% and 10.4%, respectively. A non-significant reduction in overall DE was observed in the post- compared to pre-implementation period (16.5% vs. 20.9%, OR 0.84 [0.48, 1.47], p=0.54), primarily in the high-risk cohort (15.6% vs 21.9%, OR 0.76 [0.43, 1.34], p=0.34) but not the low-risk cohort.
Conclusions: Our results suggest that harmful and non-harmful DEs occur in up to 1 of 5 hospitalized patients cared for by hospitalists. Implementation of an EHR-integrated intervention during the COVID-19 pandemic led to clinically meaningful but non-significant reductions in estimated DE rates in high-risk cohorts. The intervention components, implementation materials, and structured DE case reviews conducted by hospitalists merit further evaluation at our and other institution(s).
