NLP-enabled Decision Support for Cervical Cancer Screening and Surveillance

Cervical cancer is a preventable cause of mortality among young women. Over half of women diagnosed with cervical cancer are inadequately screened, and lack of appropriate follow-up for abnormal screening contributes to 12 percent of cancer cases. Evidence-based guidelines for cervical cancer prevention, however, are complex and require the consideration of patient-based factors generally documented only in free-text clinical notes. As such, guidelines can be challenging for healthcare providers to follow and may lead to patients receiving suboptimal preventive care. While clinical decision support (CDS) systems may improve care delivery, current CDS systems--incapable of capturing free-text data--only identify patients overdue for routine screening, do not suggest optimal screening intervals for patients not overdue for screening, and provide no assistance in the surveillance of patients with abnormal screening results.

Using a natural language processing (NLP)-enabled CDS system to capture both discrete and free-text information in a patient’s record may vastly improve a system’s ability to prompt providers in correct screening and surveillance intervals for prevention of cervical cancer. This project implemented and optimized a previously developed NLP-enabled CDS system, applied cervical cancer screening and surveillance guidelines to it, and computed the optimal recommendations for follow-up care.

The specific aims of the project were as follows:

• Develop, validate, and optimize the CDS system in the clinical setting. 
• Determine the impact of reminders to non-adherent high-risk patients. 
• Determine the impact of CDS alerts to healthcare providers. 

Researchers used national guidelines to develop a CDS system with 13 pathways for screening and 41 pathways for surveillance, and developed and tested it for accuracy. The study team used big data infrastructure to process patient records, delivered a list of patients who were overdue for surveillance to providers, and subsequently developed the CDS system workflow to deliver real-time screening and surveillance to the providers. Investigators improved the accuracy of the CDS system from 94 percent to 100 percent with iterative development and stratified testing. They also found that sending reminders to non-adherent high-risk patients significantly improved surveillance rates from 5.7 percent to 23.7 percent. Lastly, the study successfully demonstrated improvement in overall screening rates with real-time CDS to the providers.