Discussion
Recent clinical and experimental studies have underlined the TH in CM-I due to embryological developmental anomalies in the occipital bone and a hypoplastic PCF (5,6,7,18,19). However, tonsillar herniation as a single radiological diagnostic criterion for CM-I is insufficient and more parameters are needed to prevent under or misdiagnosis. A reported study by Alperin et al. identified four measures that distinguished CM-I patients from a healthy cohort with 97 percent sensitivity and 100 percent specificity (4). Another study produced a probability predictor based on a logistic regression (LR) model that took into account seven PCF measures and had a sensitivity of 93% and specificity of 92% in distinguishing patients with classic CM-I from those with a standard PCF (16).
Many studies have searched for alternatives or adjunct diagnostic radiological parameters for CM-I (4,5,16,20-23). In our study, a large CM-I population was compared with a sex-age matched control group using 5 different supervised ML programs. The primary aim was to detect the test with the most accurate result and minimal parameters.
Machine learning is a potential diagnostic tool that may be used in many disorders. The quality of each classifier was tested using three-fold cross-validation. Also, calculation of the variable importance for the features can be divided into model-based and model-independent approaches. The benefit of using a model-based method is that it is more closely related to the performance of the model and that the association structure between the predictors can be integrated into the estimation of significance. Based on the advantages of model-based methods, we used this approach for calculating the variable importance in the current study.
It has been shown that cross-validation techniques significantly improve the outcome of the models by helping to choose the best coefficients taken into consideration for the model (24). To boost the efficiency of the models, k-fold cross-validation is employed in the training stage. In the present study, three-fold cross-validation was utulized for assessing the predictive power of the supervised ML models. The random search optimization approach deals with techniques for a given problem to achieve the ’most favorable’ solution and selects the most relevant features with the output variable.
Bagged CART (F4&F11) and Random Forest (F2&F11) methods have revealed a success rate of 98% using two different parameters. This ratio increased to 100% when a third parameter was included in the analysis. Also, the diagnostic ratio of the XGBoost method was 99% using three parameters (F2, F4, and F6) but increased to 100% when a fourth parameter was included. These results reflect the usefulness and accuracy of such ML systems. The Random Forest method is found to be the most accurate test using the parameters of distance from the pons to FM (F2); fastigium to FM (F4), and Clivus canal angle (F11). This outcome supports the latest report by Urbizu et al. but with augmented accuracy (16). The high diagnostic ratio of our methods seems to be due to the matched control group and the use of advanced ML algorithms.
It is known that the posterior basicranium undergoes retroflexion during the fetal phase, pushing the basion upward and dorsally and decreasing the PCF ventral depth (25). Any developmental problem during this phase, as the spheno-occipital synchondrosis closes, may impede the normal growth of the basilar portion of the occipital bone until the 2nd decade of life (25). In our study, the distance from Fastigum to FM (F4) was the standard parameter that was used and obtained higher accuracy. This result may support the suggested role of the decreased ventral depth of PCF and hypoplastic PCF in CM-I etiology (2,5,6,7,18,21,25) .
It was suggested that the basion and the dens might grow in the embryo jointly (sclerotome resegmentation), supporting the theory that a shift in the location of the basion could result in a concomitant change in the dens position, leading to more craniocervical changes in these patients (26). This relation was concluded as a necessity to include the Odontoid in such measurements (16). In our study, the Odontoid was included as a parameter but provided different values of accuracy that in Random Forest test no change was observed; but in Bagged CART test, the value of accuracy was increased and decreased in different combinations. Also, inclusion of clival and Corpus Callosum measurements was found to increase the classification accuracy. The crowded PCF and CSF dynamics may decrease the reliability of mobile points like pons when compared with osseous structures like Clivus, dorsum sella and odontoid. These findings should be evaluated in cooperation with literature-based data that support the role of crowded PCF in the etiology of CM-I (5,27,28,29).
The optimal surgical management of CM-I is still controversial. Posterior fossa decompression with or without duroplasty is the most recommended surgery (28, 30-32). However, our results may guide further studies that anteriorly localized structures seem to influence the PCF and etiology of CM-I as much as posterior ones. This may also aid different surgical approaches and management ways.
Regarding the limitation of the study, even statistically matched, the dominance of female cases in the study population might be considered as a minor bias. A larger and more homogenous study population may be more appropriate. Also, TH was calculated using midsagittal MR images, excluding any possible asymmetry regarding parasagittal structures that may influence symptomatology. Larger studies may be designed in order to exclude any possible difference in race and regional factors that may have an effect on morphometric anatomy.