The Development and Validation of an Electronic Health Record (EHR)-Based Algorithm for Identifying Treatment Failure in Newly Diagnosed Hodgkin Lymphoma (HL) Treated in a US Community Oncology Setting

Introduction

Most existing cancer registries do not fully capture relapse or progression of cancer. Further, clinical trials for upfront treatment typically withdraw patients from study at the time of treatment failure and only collect minimal information about approaches to and outcomes with salvage therapy. As such, patient outcomes other than mortality often are available only through formal chart review, which is resource intensive. For HL, although many patients are successfully cured, a sizeable subset of patients experience treatment failure. Most of the data describing failure arises from prospective clinical trials, despite most patients receiving treatment in the community setting. Therefore, data from the community setting are needed to understand the experience of real-world patients and to inform and improve the quality of care in oncology practices. To facilitate observational research and quality improvement projects on HL treatment failure, we examined an EHR-based algorithm for identifying treatment failure in HL.

Methods

All HL patients who initially presented with advanced stages II-IV disease between 2007-2012 at Kaiser Permanente Southern California (KPSC) were identified using KPSC's cancer registry. An algorithm for identifying treatment failure was proposed by study oncologists. Treatment failure was defined as relapse after complete remission of > 60 days' duration or disease progression on active treatment. The algorithm was defined as meeting one of the following: (1) Initiation of any second chemotherapy regimen, including use of brentuximab vedotin, or PD-1 inhibitor beyond initial regimen; or (2) stem cell transplantation (SCT). Data needed to construct the algorithm were collected from KPSC's EHR, including the chemotherapy database and transplant registry. The algorithm was applied to all HL cases up to 4 years after diagnosis, as most HL relapses occur within this timeframe. Manual chart reviews were performed for all potential treatment failure cases identified by the algorithm, and for one-third of randomly-selected HL cases not captured by the algorithm. Each case was abstracted by trained study staffed and reviewed for quality control by two study oncologists.

Results

Of the 470 HL patients newly diagnosed during the study period (91% diagnosed at age 18 and older, 57% male), the algorithm identified 80 as potential treatment failure cases. Chart review confirmed 75 of 80 as actual treatment failure. The algorithm misclassified relapse for 5 cases who had emergence of a second malignancy for which therapy was instituted after the completion of initial therapy for HL. The positive predictive value (PPV) was 94% (95% confidence interval: 89%-99%). The negative predictive value (NPV) based on the review of 127 patients not captured by the algorithm was 100%. Based on the high PPV and NPV, our review suggested both a high sensitivity and specificity of the algorithm.

Conclusion

This EHR-based algorithm, based on receipt of second-line therapy beyond initial treatment and SCT, was highly effective in capturing relapse or disease progression cases among HL patients in this integrated health care delivery system. Application of this algorithm will allow for better long-term follow up for novel agents in observational studies, which is critical for understanding the cost-effectiveness of therapeutic agents in real-world patients. External validation using data from other health care settings or sources such as claims data will be useful to further confirm the generalizability of this algorithm. Altogether, this algorithm could assist practices in the identification and prospective tracking of the relapsed/refractory HL patient population for clinical and non-clinical outcomes, such as cost, health-related quality of life, and sequelae of treatment.