A Novel Dashboard to Improve Lung Cancer Detection
Project Lead: Christopher Grondin, MD
Project AIM(s): We will reduce the number of patients with pulmonary nodules identified in our emergency department that do not receive appropriate follow-up imaging by 50% by March 2021.
Narrative Description
This intervention bridges the communications gap between VA patients that have pulmonary nodules incidentally detected through imaging obtained for other reasons and primary care providers. This aims to help physicians coordinate further imaging and sometimes biopsy to detect lung cancer earlier. Specifically, the aim is to reduce the number of patients with pulmonary nodules identified in our emergency department that do not receive appropriate follow-up imaging by 50% over the period of the yearlong project.
The team of 3 MDs manually reviewed patient CT chest exam data to determine if the patient should be flagged for guideline directed (Fleischner) follow up. A data analyst used the identified patients’ data to train a natural language processing (NLP) algorithm (SimpleNLP) to automate identification of patients in need of follow up. The initial run of the model on a sample of patients revealed sensitivity of 52% and specificity of 77%. To increase sensitivity/specificity, the team analyzed false positives/negatives and corrected the model. Using this data, the team created a dashboard to track patients requiring follow-up which will automate the creation of provider notes and patient letters to increase guideline-directed follow-up.
In the pilot testing the model and dashboard, the team searched all CT chest scans that had been done on patients of one of the co-investigators with a partial appointment in primary care over the last three years. This returned 86 scans on 48 unique patients. 11/12 of these patients were flagged appropriately with revised algorithm. One patient had a 6-8mm nodule that merited overdue follow-up. The team met with VA patient panel and VA Ann Arbor primary care physicians who provided insight into the cadence of desired alerts from the dashboard as well as wording of patient notifications.
Further testing of the model is planned for a separate group of 500 patients to obtain updated sensitivity/specificity with plans to expand pilot to all VAs. This project can be reproduced with other populations and at other centers. The natural language processing algorithm should undergo repeat validation at a new center as there are differences in the format of imaging reports. The basic premise of identifying patients with incidental lung nodules and alerting their primary care providers to ensure guideline directed follow-up is readily repeatable.
Diagnostic quality problem type, failure, or category (symptoms, observed problems, gaps in performance) addressed by the intervention
- Information gathering
- Information integration
- Information interpretation
- Establishing an explanation (diagnosis)
- Communicating the explanation to the patient
Root causes/causative factors addressed by the intervention
- Workflow (includes testing, follow-up, and referrals)
- Information sharing and accessibility
Setting of the diagnostic quality improvement intervention
- Ambulatory medical care setting (e.g., clinic, office, urgent care)
- Emergency department
- Radiology/imaging