Comparing the Documented Pressure Injury in MIMIC-III: An “UpSet” Visualization

Background/significance: Pressure injury is one of the key patient safety and care quality indicators. To monitor PI prevalence and explore risk factors, electronic health records (EHRs), have been adopted in PI research. However, a recent NLM study found great discrepancies in both prevalence and cases with a Venn diagram comparing the identified PI cases in the diagnosis, stage data in the chart events, and clinical notes in MIMIC III, a freely accessible critical care database.

Purpose: Since other available PI features such as site in the chart events and the number of PI on each patient have not been compared, this study thus aims to further and more comprehensively compare all identified PI cases based on diagnosis codes, chart events, clinical notes, and procedures in MIMIC-III.

Methods: ICD-9 diagnosis codes, chart event item IDs, keywords, and CPT numbers for PI were used to extract PI datasets. As Venn diagram gets complex when visualizing over 4 sets, we applied UpSetR to generate static UpSet plots to visualize and compare the identified PI set interactions based on the ICD-9 diagnosis codes, chart events, clinical notes, and CPT events. The numbers of PI site and stage were analyzed and then also compared to see if they were the same among patients with diagnosis or chart event data.

Results: 32,211 patients in MIMIC III with either of the following data were included: 1) diagnoses on PI stage or site within up to #39 total diagnoses; 2) chart events on Braden risk sub or total scale score, #1 to #3 PI’s site, stage, depth, drainage, width, cleansing, treatment, wound base, odor, pressure support, pressure reduce device, heal, amount (drainage), or length items from one system CareVue or up to #10 stages from MetaVision, the other system; 3) detected and non-negative PI clinical notes; and 4) procedures of wound debridement or wet-to-dry dressings. PI documented in the MetaVision chart event system and CPT events were incomplete. The number of patients with PI were 1837, 2850 and 6994 respectively based on the diagnosis, chart events, and clinical notes.

UpSet visualizations presented the great discrepancies in PI documentation: (1) chart events captured much more PIs; (2) stage was less documented than site in the diagnosis or charts; (3) PI number was not a chart feature and inconsistent across diagnosis or chart when comparing site with stage; (4) chart events on PI depth, width, drainage amount, odor, and cleansing were less documented when compared with other features in the chart events. The differences in the number of PI (site – stage) ranged from -2 to 4 in the diagnosis and from -3 to 2 in the chart events. For patients with both site and stage data, 810 patients (44.1%) reported the same number of site and stage in the diagnosis and 2,211 (98.0%) in the charts.

Conclusions/implications: PI documentation needs improvement. Upset plots could be used as clinical informatics tools to inform documentation quality. PI research may use EHR chart event data and needs to validate the results. 

Purpose: Since other available PI features such as site in the chart events and the number of PI on each patient have not been compared, this study thus aims to further and more comprehensively compare all identified PI cases based on diagnosis codes, chart events, clinical notes, and procedures in MIMIC-III.

Methods: ICD-9 diagnosis codes, chart event item IDs, keywords, and CPT numbers for PI were used to extract PI datasets. As Venn diagram gets complex when visualizing over 4 sets, we applied UpSetR to generate static UpSet plots to visualize and compare the identified PI set interactions based on the ICD-9 diagnosis codes, chart events, clinical notes, and CPT events. The numbers of PI site and stage were analyzed and then also compared to see if they were the same among patients with diagnosis or chart event data.

Results: 32,211 patients in MIMIC III with either of the following data were included: 1) diagnoses on PI stage or site within up to #39 total diagnoses; 2) chart events on Braden risk sub or total scale score, #1 to #3 PI’s site, stage, depth, drainage, width, cleansing, treatment, wound base, odor, pressure support, pressure reduce device, heal, amount (drainage), or length items from one system CareVue or up to #10 stages from MetaVision, the other system; 3) detected and non-negative PI clinical notes; and 4) procedures of wound debridement or wet-to-dry dressings. PI documented in the MetaVision chart event system and CPT events were incomplete. The number of patients with PI were 1837, 2850 and 6994 respectively based on the diagnosis, chart events, and clinical notes.

UpSet visualizations presented the great discrepancies in PI documentation: (1) chart events captured much more PIs; (2) stage was less documented than site in the diagnosis or charts; (3) PI number was not a chart feature and inconsistent across diagnosis or chart when comparing site with stage; (4) chart events on PI depth, width, drainage amount, odor, and cleansing were less documented when compared with other features in the chart events. The differences in the number of PI (site – stage) ranged from -2 to 4 in the diagnosis and from -3 to 2 in the chart events. For patients with both site and stage data, 810 patients (44.1%) reported the same number of site and stage in the diagnosis and 2,211 (98.0%) in the charts.

Learning Outcome: After completing this learning activity, the participant will be able to assess innovations being used by other professionals in the specialty and evaluate the potential of implementing the improvements into practice.