Development and validation of coding algorithms to identify patients with incident lung cancer in United States healthcare claims data

Abstract

Purpose: Our aim was to develop and validate a practical US healthcare claims algorithm for identifying incident lung cancer that improves on positive predictive value (PPV) and sensitivity observed in past studies. Methods: Patients newly diagnosed with lung cancer in Surveillance, Epidemiology, and End Results (SEER) (gold standard) were linked with Medicare claims. A 5% Medicare "other cancer" sample and noncancer sample served as controls. A split-sample validation approach was used. Rules-based, regression, and machine learning models for developing algorithms were explored. Algorithms were developed in the model building subset. Rules-based algorithms and those with the highest F scores were evaluated in the validation subset. F scores were compared for 1000 bootstrap samples. Misclassification was evaluated by calculating the odds of selection by the algorithm among true positives and true negatives. Results: A practical single-score algorithm derived from a logistic regression model had sensitivity = 78.22% and PPV = 78.50% (F score: 78.36). The algorithm was most likely to misclassify older patients (ages ≥80 years) or with missing data in the SEER registry, shorter follow-up time in Medicare (<3 months), insurance through Veterans Affairs, >1 cancer in SEER, or certain Charlson comorbidities (dementia, chronic pulmonary disease, liver disease, or myocardial infarction). Conclusion: In this dataset, a practical point-based algorithm for identifying incident lung cancer demonstrated significant and substantial improvement (7.9% and 23.9% absolute improvement in sensitivity and PPV, respectively) compared with a current standard.

Algorithm to Identify Incident Epithelial Ovarian Cancer Cases Using Claims Data

“…However, algorithms to identify incident cases exist for many other cancer types. [16][17][18][19][20][21][22][23] Our algorithm is similar to proposed by Nattinger et al 16 to identify incident cases of breast cancer, which likewise used an initial diagnosis code to identify breast cancer, a secondary step to identify cases with high likelihood of being true breast cancer (using procedural codes instead of chemotherapy and surgical claims), excluded nonbreast cancer cases, and removed prevalent cancer cases. Other algorithms to identify incident cancer cases use logistic regression models to derive predictor variables 19,21 or require various elements of inpatient and outpatient diagnosis codes, procedures, and/or laboratory values within defined time points.…”

“…In particular, there is the potential for inaccurate or missing data and a paucity of validated clinical and pathologic data, which creates challenges for accurate cohort selection for research purposes. 5,14,15 Although previous studies have developed algorithms using claims data to identify cases of multiple other types of cancers, [16][17][18][19][20][21][22][23][24] currently, there is no algorithm to identify incident ovarian cancer cases. As research into ovarian cancer using administrative databases continues, it is imperative that these cases can be correctly ascertained from claims databases to make accurate research inferences and conclusions.…”

Development and Validation of Coding Algorithms to Identify Patients with Incident Non-Small Cell Lung Cancer in United States Healthcare Claims Data

“…Full details of the data source used in this study were published previously. 15 The protocol was reviewed and considered exempt by the Quorum Review institutional review board prior to approval by the National Cancer Institute for SEER-Medicare data use.…”

“…We searched for candidate variables in the pre- and post-index periods in Medicare claims. Based on the algorithms that performed best in our previous study, 15 the following models were explored in this study: a single logistic regression model, a single logistic regression model with interactions, gradient boosting, and neural networks. The multilayer perceptron neural network had two hidden layers in the network and two neurons in each hidden layer because five hidden nodes did not improve the F-score in the model building set.…”

“…We re-categorized some of these individuals as lung cancer cases, similar to that done by Nattinger and colleagues, 16 which resulted in 193 patients being re-categorized from non-cancer to lung cancer in this study. Although these patients were included as lung cancer cases in the previous study, 15 they were excluded from this study. This was necessary since they were not in the SEER registry and did not have information on lung cancer type.…”