Identifying Potential Associations Between Patient Reported Outcomes and Normal Structure Dose Metrics in Head and Neck Cancer Patients Using a Big Data Approach

Abstract

Including patient reported outcomes (PROs) as data elements automatically aggregated into Big Data Analytics Resource Systems (BDARS) generates vast amounts of data linking potential associations of PRO responses to other key data elements. Extracting clinically useful insights from these associations helps demonstrate value in collection of PROs. A challenge in analysis of the large volume of potential associations is development of statistical filters identifying associations that are both clinically and statistically significant. In this study, we examined associations between PROs and normal tissue dose measures from head and neck cancer patients, then applied statistical filters to identify those most likely to be clinically relevant. Two hundred sixty-two head and neck cancer patients previously treated in our department, who had completed PRO questionnaires at follow-up appointments, and had available normal tissue dose measures were included in the study. The PRO questionnaires included the University of Washington QOL Questionnaire (UW-QOL), University of Michigan Xerostomia Questionnaire (XQ), and pain, fatigue and overall quality of life components from the Linear Analogue Self-Assessment scale (LASA-3). Using a big data approach, we analyzed interactions between each of the individual PROs from these questionnaires and mean and maximum doses from all pertinent normal structures. For each combination of PRO and dose measure, statistical metrics (AUC, PPV, NPV, sensitivity, p-value) were generated using all possible PRO response cutoffs to separate positive from negative outcomes. Filters for each of these metrics were applied individually and in combination, and the remaining number of PRO responses and normal structure dose pairs for a given cutoff (result pairs) at each iteration were calculated. A total of 4680 result pairs were identified for both mean and maximum dose. Using an AUC filter of 0.65 and dose cutoffs of 24 Gy for mean dose and 45 Gy for max dose returned 82 result pairs for mean dose and 69 result pairs for max dose. Additional filters including sensitivity of 0.50, positive predictive value of 0.1, negative predictive value of 0.5, and p-value of 0.02 reduced the remaining result pairs to 20 and 25, respectively, for mean dose and maximum dose. Of this total of 45 result pairs, 22 involved submandibular gland doses and one involved parotid gland dose. Associations between brainstem dose measures and activity and overall quality of life were also noted, although only for the lowest PRO response cutoffs. Applying statistical metric filters to all combinations of PRO responses and normal structure dose measure pairs identified a manageable group of result pairs for further investigation. In this hypothesis-generating study, a stronger association was noted between PROs and submandibular gland doses than parotid gland doses in patient treated for head and heck cancers.