Automated Population of Oncologic Clinical History From Electronic Medical Records in PET/CT Reports

Introduction: The clinical information(CI) section is important to the PET/CT report. In oncology, this includes the cancer diagnosis, relevant treatment, and clinical question. The nuclear medicine(NM) physician collects this information from multiple, often disparate, sites, including order forms, patient questionnaires, and electronic medical records(EMR). This must be done with each patient and can be a time burden. Given the increasing demand of NM workload, this task may be skipped or skimmed due to a busy reading schedule, despite studies showing that the context from a comprehensive CI section improves interpretation accuracy, clinical relevance, and reporting confidence. This also must be done with each new report as disease management is constantly evolving. While they may want to, most physicians do not trust the clinical information from the previous report of another colleague. This makes both time and trust issues in establishing a comprehensive clinical context. We set out to create a system that addresses both concerns. NM reporting templates are becoming more structured with the ability to auto-populate information. Report automation saves time, improves accuracy, and decreases fatigue. Therefore, we built the infrastructure to successfully auto-populate the CI into an oncology PET/CT report from the EMR at a major cancer center.

Methods: We first established that diagnosis, biopsy, medical therapy, radiation therapy, surgery, and their dates were the requisite elements of the CI. We use a structured layout template for PET/CT reports with a form field for the CI section. We decided to present the requisite elements as bulleted line items from oldest to newest within this field in the following format, * Date: Element (Additional Information). We then created a software application for manual data entry of these requisite elements from the coded information in the EMR (image 1, 2). A database of RadLex terms and their codes was created, allowing only structured pick menu entry of all requisite elements, image 3. For example, in image 4, the diagnosis is "breast cancer RID45682". While the date and element are structured information from the EMR, the additional information component allowed for free text entry to add detail. For example, in the case of medication this component would allow you to add the specific drug name, image 5. Data entry was performed by a Radiology Assistant under training by a NM attending with periodic QA/QI. The information for each patient was stored and visualized by a Tableau dashboard, Image 6. And this information would auto-populate in the CI field when a reader opened the study in Powerscribe, Image 7.

Results: We successfully created a software application that allows manual input of coded EMR CI in the form of RadLex terms. We built the infrastructure to store these data, visualize them via a Tableau dashboard, and auto-populate them into structured PET/CT reports. Clinical information for over 1000 patients was created, and QA/QI was performed on approximately 10%, leading to the evolution of format over time. User feedback from attendings who reported the scans was gathered, with comments on style, structure and quantity. The most relevant criticism was, given that many of the patients at our center have repeated follow-up imaging over a long course, unedited addition of subsequent treatments would lead to an increasingly long CI field that would be difficult to read and unmanageable in size. Therefore, the final version only includes initial diagnosis and limited most recent treatments for each scan time point. Since the EMR information is coded there is potential for automation of data input via machine learning, which represents the future goal of this project.

Conclusions: We built the infrastructure to successfully auto-populate the clinical information into an oncology PET/CT report from the electronic medical record at a major cancer center, with the future goal of automation by machine learning.

• Download figure
• Open in new tab
• Download powerpoint