Study: Multilevel Data Model Helps Estimate Likelihood of Delays in Cancer Treatment

Machine learning models in multilevel data sources can help to estimate the likelihood of delays in starting cancer treatments, according to the results of a study published in JAMA Network Open.¹

Image credit: peterschreiber.media – stock.adobe.com

In the study, investigators incorporated data from electronic health records (EHR) and social determinants of health (SDoH) into the model to determine the likelihood in delays when starting cancer therapy.¹

The impact of a delay in treatment can increase the risk of mortality for those with cancer. A 4-week delay can increase the risk of mortality across surgical, systemic treatment, and radiotherapy indication for at least 7 cancers, according to research published in The BMJ.² Additionally, the association between delay and mortality risk was significant for 13 of the 17 indications included in the study.²

Because of this association, the authors of the current study aimed to determine and validate a model of learning to estimate the probability of treatment delay with multilevel data sources.¹

The authors of the study evaluated 4 different machine learning approaches in the cohort study for the likelihood of treatment delays greater than 60 days. This included: least absolute shrinkage and selection operator (LASSO); Bayesian additive regression tree; gradient booting; and random forest.¹

Discrimination, calibration, and interpretability were criteria that the investigators considered. The data set included clinical, demographic, and neighborhood-level census information derived from EHR, cancer registry, and the American Community Survey. Furthermore, patients with invasive breast, lung, colorectal, bladder, or kidney cancers who were diagnosed from 2013 to 2019 and treated at a comprehensive cancer center were included. Investigators analyzed the data from January 2022 to June 2023.¹

The study authors used the area under the receiver operating characteristic curve (AUC-ROC) as the primary metric to evaluate the model’s performance.¹

A total of 6409 individuals were included in the study, of whom 2576 had breast cancer, 1738 had lung cancer, and 1059 had kidney cancer. Approximately 25.3% of the individuals experienced a delay in treatment greater than 60 days. For the LASSO model, there was an AUC-ROC of 0.713.¹

Investigators found a lower likelihood of delay when the diagnosis was done at the treating institution; when it was the first malignant neoplasm detected; if the individuals were Asian, Pacific Islander, or White; had private insurance; or lacked comorbidities. There was a greater likelihood of delay in the extremes of neighborhood deprivation.¹

Additionally, investigators found that the model’s performance was lower for Black individuals, those with race and ethnicity that were not non-Hispanic White, and those living in the most disadvantaged neighborhoods.¹

However, the study authors did find that the performance was similar when the neighborhood SDoH was considered and left out. This was true for the overall study population as well as across the sociodemographic subgroup populations.¹

The model’s performance was similar in individuals who were diagnosed internally and externally.¹ The study authors determined that future work should find additional ways to use SDoH data to help improve the model’s performance, especially in more vulnerable populations.¹

Investigators cited a possible limitation as variables excluded from the EHR data, which could have affected access and timing of cancer therapies.¹

References

1. Frosch ZAK, Hasler J, Handorf E, DuBois T, et al. Development of a multilevel model to identify patients at risk for delay in starting cancer treatment. JAMA Netw Open. 2023;6(8):e2328712. doi:10.1001/jamanetworkopen.2023.28712
2. Hanna TP, King WD, Thibodeau S, Jalink M, et al. Mortality due to cancer treatment delay: systematic review and meta-analysis. BMJ. 2020;371:m4087. doi:10.1136/bmj.m4087