Associations Between First Year and Early-Year Costs in Pediatric Spina Bifida Patients - A Longitudinal Machine-Learning Based Analysis
Venkat M. Ramakrishnan, M.D., Ph.D.1, Michael Li, M.B.A.2, Carlos Estrada, M.D., M.B.A.3, Hsin Hsiao S. Wang, M.D., M.P.H., M.B.An3.
1Brigham and Women's Hospital, Boston, MA, USA, 2Massachusetts Institute of Technology, Boston, MA, USA, 3Boston Children's Hospital, Boston, MA, USA.
BACKGROUND: Spina bifida (SB) is a relatively common, chronic multi-system disorder that can present heterogeneously. SB patients often form lifelong relationships with healthcare providers given their need for long-term functional monitoring and, with severe disease, extensive medical and surgical interventions. Therefore, the disease carries a likely potential for high long-term costs, placing financial and resource strains on patients, providers, and healthcare systems. Our goal was to assess if machine learning (ML) algorithms could extrapolate early-year costs (EYCs) from first-year costs (FYCs) such that longer-term patient needs could conceivably be determined from their needs during the first year of life. METHODS: The Aetna database was used to access patient information from 2006-2019. The relevant ICD-9/-10 diagnosis codes and CPT codes were used to identify pediatric patients with SB status post initial myelomeningocele closure. Those included were required to have complete, continuous records from birth until age 5. Exclusion criteria included lapses in Aetna coverage and death before age 5. FYC and EYCs were determined using a K-nearest neighbors algorithm, with cluster number determined by within-group sum of squares. The predictive performance of the model was then assessed via accuracy, specificity, and sensitivity. Results are reported as means with 95% CI, and all analysis was performed using R. RESULTS: The Aetna database initially yielded 351 patients, of which 58 underwent further analysis after exclusion criteria were applied. From this, two clusters (high-cost group (HCG) and low-cost group (LCG)) were identified. Excluding surgical cost, the HCG demonstrated an average 5-year expenditure of $327,000, while the LCG spent $45,000. The median FYC was $22,000, serving as the threshold for extrapolative EYC analysis. Actual annual mean EYCs were found to be $55,546 for the HCG (aggregate mean = $222,183) and $11,138 for the LCG (aggregate mean = $44,550). When the ML algorithm was applied to FYC, the algorithm identified 24/29 HCG and 26/29 LCG members correctly, yielding a predictive accuracy of 0.86 (95% CI: 0.75-0.94), sensitivity of 0.89, and specificity of 0.83. Moreover, HCG patients were found to have increased facility, pharmacy, and service needs. CONCLUSIONS: This is the first study to predict long-term costs in the SB population using a complete database. The ML algorithm is useful in identifying high-risk patients with increased disease severity who may require additional resources. This may offer benefit with respect to financial planning, resource allocation, policy making, and education.
Back to 2022 Abstracts