Akademik Veri Yönetim Sistemi
Study design: retrospective analysis of prospectively collected data. Objective: to investigate whether two clustering approaches applied to the same database would lead to differences in the minimal clinical important difference (MCID) for health-related quality of life parameters (HRQoL). Summary of Background Data: Machine learning approaches are being increasingly employed for the analysis of complex and heterogeneous settings such as that of adult spine deformity (ASD). However, it is not yet clear whether and how the choice of number and type of variables impacts the outcomes of a study. Methods: Two previously published clustering approaches (C12 and C16) were applied to a multicentric database of ASD patients who underwent surgery and had a minimum follow-up of one year. After clustering, the MCID for the Oswestry Disability Index, SRS-22, and SF-36 PCS were calculated for all clusters using the ROC method. Results: Data from 516 patients were available. Both algorithms led to a division of the database in three clusters, which presented similar characteristics both for C12 and C16. In particular, patients in clusters 1 to 3 presented an increasing level of imbalance and disability. The MCID for ODI, SRS-22, and SF-36 for each cluster differed between C12 and C16, but a similar pattern of increase of the MCID from Cluster 1 to Cluster 3 was observed for all HRQoL parameters and in both C12 and C16. The error rate, however, was smaller for C16. Conclusion: Different clustering algorithms applied to the same database allowed to obtain similar clusters of ASD patients. However, the obtained MCIDs for the evaluated HRQoL parameters were different, highlighting the relevance of the choice of variables for the investigation of these parameters. The results suggest that clinically-driven clusters should be used when investigating clinical outcomes, as they allow for a smaller error rate.
