Search for a command to run...
To the Editor: In our recent publication, we identified 3 presenting phenotypes of Chiari type 1 malformation and syringomyelia using a clustering approach that integrated clinically informed and data-driven feature selection.1 As part of our methodology, we applied the Laplacian score, an unsupervised ranking method, to select informative features for clustering.2 In our original analysis, we selected features based on literature that interpreted higher Laplacian scores as indicating more informative features.3-5 In subsequent work, we have encountered alternative interpretations in the literature supporting lower scores as more desirable.6,7 On careful investigation, we believe the latter offers a more appropriate interpretation. The Laplacian score quantifies how much a feature varies across locally neighboring data points. Features with lower scores demonstrate greater locality-preserving power, as they vary less within local neighborhoods, thus better maintaining the intrinsic structure of the data set. Given the presence of conflicting interpretations in the literature, we believe it is important to raise awareness of this methodological ambiguity and to provide results with feature selection based on lower Laplacian scores. As such, we have repeated our analysis using the revised criterion and summarized the updated findings in comparison with the original results. The updated results are presented in Figure, with details provided in Supplemental Digital Content eTable 1 and eFigures 1 and 2 (https://links.lww.com/NEU/F285). Although this adjustment changed the ranking of data-driven features, many of the newly selected features overlapped with those from the original analysis, as shown in the Supplemental Digital Content Methods (https://links.lww.com/NEU/F285). This consistency is attributable to the integration of clinical guidance, which involved clinical surveys and grouping features into clinically meaningful categories.FIGURE.: Top panel shows results from the original analysis, and bottom panel shows updated results using the revised Laplacian score criterion.As shown in Figure, the clustering patterns and clinical interpretations continue to support the primary conclusions of the study, with the data-driven methodology supporting the inclusion of specific characteristics identified as important by clinical experts, including syrinx diameter, presence of occipital/suboccipital headaches, and degree of tonsillar descent. The revised methodology also supports the inclusion of other clinical or radiologic data including reflexes, sensory changes, and pBC2 distance. The cluster characteristics and overall phenotypes remain largely the same, particularly with respect to defining factors such as diagnosis age, syrinx size, bulbar symptoms, presence of hydrocephalus, spinal column abnormalities, and degree of tonsillar descent. Interestingly, the distribution pattern of several other characteristics such as physical exam findings and presence other medical issues, many of which had lower magnitude differences in the initial analysis, has changed slightly within each group. Although this may suggest that these factors are less important for clustering as they are not preserved between analyses, the exact implications remain unclear, and further investigation into the complex relationship of these factors with Chiari type 1 phenotypes is certainly warranted. We appreciate the opportunity to share this updated analysis.