No significant differences in the degree of muscle infiltration were demonstrated between the same muscles of both extremities, after correction for multiple comparisons (Table 1)

Table 1: Asymmetry of the degree of fatty infiltration between paired muscles. The column ‘statistic’ corresponds to the Wilcoxon test statistic. The difference corresponds to the subtraction between the degree of fatty infiltration of the right and left side. A p-value of 0.001 (0.05 corrected for 35 multiple comparisons) was considered significant.

Fig 2: Degree of fatty infiltration for each muscle ordered by body segment in patients with dysferlinopathy. Violin plots represent data dispersion, mean, and standard deviation. Each pair of adjacent muscles corresponds to right and left-sided muscles respectively.

The pattern of fatty infiltration in dysferlinopathy

Figure 3 shows a heatmap of the degree of fatty infiltration for patients with dysferlinopathy. Some muscles present a greater degree of fatty infiltration than others. For example, a lower degree of fatty infiltration is observed in the pelvic muscles, and a higher degree of infiltration in the thigh and leg muscles, with relative respect for the sartorius and gracillis, as well as the popliteus, finger flexors, and posterior tibialis.

Fig 3: Heatmap of the degree of fatty infiltration of 70 muscles in patients with dysferlinopathy. Each column corresponds to a muscle and each row to a patient. The degree of infiltration according to the Mercuri scale for each muscle is represented by the colour intensity from 0 (blue colour) to 4 (yellow colour), according to the adjacent colour bar. A mean imputation of each column was used for missing values.

Figure 4 shows a violin graph in which the previous results are corroborated. Moreover, it is observed that for severely affected muscles, such as soleus, semimembranosus, gastrocnemius, peroneus, flexor digitorum longus, and biceps, corresponding to the posterior compartments of the leg and thigh respectively, there is a low dispersion of the data. Similarly, some of the relatively respected muscles, such as the piriformis, popliteus, and gracilis, present a lower degree of variability; that is, they remain relatively respected in most patients. However, other muscles that present on average a low value of fatty infiltration, such as the psoas iliacus and sartorius, present a high variability, demonstrating greater heterogeneity.

Fig 4: The average degree of fatty infiltration of muscles in patients with dysferlinopathies is in ascending order by degree of involvement. Each violin plot represents the mean, standard deviation, and distribution of the data.

Statistically significant differences were observed between the gluteus minimus, tensor fascia latae, obturator externus, pectineus, adductor brevis, quadratus femoris and vastus medialis, lateralis and intermedius muscles; triceps suralis, peroneus, gastrocnemius and soleus, which present a markedly greater degree of fatty infiltration compared to pelvic muscles, such as psoas iliacus, gluteus maximus and gluteus medius (Figure 5). Together with the above, it is observed that the sartorius and gracilis muscles present relative respect, concerning the rest of the thigh and leg muscles; however, no significant differences were detected for the pelvic muscles, forming another group of relatively respected muscles. Another muscle that also shows statistically significant respect concerning the rest of the musculature corresponds to the popliteus, even in comparison to the other muscles of the adjacent compartments of the leg.

Fig 5: Comparison of the degree of fatty infiltration between the average of each muscle pair according to the Wilcoxon test. The p-value was adjusted for multiple comparisons. A p-value of less than 0.001 (0.05 corrected for 35 comparisons) was considered significant. The colour scale shows the difference between the degree of fatty infiltration of the first and second muscles. Each cell corresponds to the subtraction between each pair of muscles. Signs represent the p-value ♦(0.001/35), or (0.01/35), * (0.05/35).

The hierarchical clustering in Figure 6 shows that there are muscle groups that cluster in an unsupervised manner in different muscle compartments such as pelvic muscles with a lower degree of fatty infiltration. Muscles of the anterior compartment; muscles of the posterior compartment of the thigh; lateral compartment of the leg, peroneus longus and brevis muscles. However, a set of muscles with a similar degree of fatty infiltration that does not belong to the same compartment is also observed, involving the gracilis, sartorius, short-head biceps, extensor digitorum, anterior and posterior tibialis.

Fig 6: Cluster map of the degree of fatty infiltration for each patient with dysferlinopathy. Each of the 70 muscles tested is represented in each column. The degree of fatty infiltration according to the Mercuri scale is represented by the intensity of the cell colour, according to the adjacent colour bar. The dendrograms along the vertical and horizontal axis correspond to the result of Hierarchical Clustering using the ‘City Block’ metric, between patients and between muscles, respectively. Two contiguous cells in the dendrogram represent a higher degree of similarity.

The pattern of fatty infiltration in dysferlinopathy

Table 2 and Figure 7 show the B coefficients of the logistic regression, with their respective OR, CI, and p-value. It can be seen that the obturator externus (OR 18.63, 95% CI 5.32 - 65.21 ), gluteus minimus (OR 5.89 95% CI 2.4 - 14.45), psoas (OR 5.34, 95% CI 1.82 - 15.61), flexor hallucis longus (OR 5.42 95% CI 2.36 - 12.44), peroneus longus (OR 1. 2 CI 95% 1.18 - 9.35), soleus (OR 3.03 CI 95% 1.12-8.21), medial gastrocnemius (OR 2.99 CI 95% 1.29 - 6.9), flexor digitorum (OR 2.31 CI 95% 1 - 5.33), were associated with a statistically significant positive B, compared to the rest of the dystrophies. On the other hand, the quadratus femoris (OR 0.32 CI 95% 0.11 - 0.94), gracillis (OR 0.25 CI 95% 0.07 - 0.82), adductor magnus (OR 0.19 CI 95% 0.06 - 0.61), piriformis (OR 0.14 CI 95% 0.04 - 0. 47), extensor digitorum (OR 0.13 CI 95% 0.04 - 0.46) and gluteus medius (OR 0.02 CI 95% 0.0 - 0.1), presented a statistically significant negative B concerning the rest of the dystrophies.

Table 2: Coefficients of the logistic regression model adjusted for patients with dysferlinopathy versus the rest of the dystrophies.

Fig 7: Coefficients of the logistic regression model adjusted for patients with dysferlinopathy versus the rest of the dystrophies. The horizontal axis represents the Odds Ratio (OR) value for each characteristic. The vertical axis corresponds to 31 muscles used as variables after eliminating variables with high collinearity. The dots represent the OR value and the black lines are the confidence interval (CI). The blue circles correspond to the statistically significant variables of the model. A significant value of less than 0.05 was considered significant.

Figure 8 shows a confusion matrix for the predictive logistic regression model, where in the case of dysferlinopathy, the model correctly predicted 39 of the 41 patients with this dystrophy. One patient with dysferlinopathy was classified with anoctaminopathy and another as oculopharyngeal dystrophy.  

Fig 8: Confusion matrix with the prediction results of the logistic regression model on the test set. The results of the actual classes are indicated on the vertical axis and the predicted classes on the vertical axis. The values indicated on the diagonal of the matrix correspond to the true positives.

Table 3 shows the precision, recall, and f1-score metrics for each of the classes (see supplementary methods). A weighted accuracy of 0.88 and a mean accuracy of 0.79 were obtained. For dysferlinopathy, the model presented precision and recall of 0.95, reaching the highest f1 -Score of 0.95, together with facioscapulohumeral dystrophy that obtained a precision of 0.94 and recall of 0.96, with an f1-score of 0.95.

Table 3: Metrics obtained for the multi-class logistic regression model obtained on the test set for each of the dystrophies. Metrics are calculated by true positives (VP), true negatives (VN), false positives (FP) and false negatives (FN). Precision (VP / VP + FP), Recall (VP / VP + FN), F1-Score (2 x Precision x Recall / Precision + Recall). Accuracy (VP + VN / VP + VN + FP + FN). Support (number of cases in the test set for each dystrophy).

Explainability of logistic regression models

Figure 9 shows the 9 variables with the greatest impact on the decision of the model and the sum of the remaining 26 variables. Having greater fatty infiltration of the obturator externus, flexor hallucis longus, and soleus has a positive impact on the model for classifying a patient with dysferlinopathy. On the other hand, having low fatty infiltration in the gluteus medius, gracilis, piriformis, adductor magnus and long head of the biceps have a positive impact on the classification of a patient as having dysferlinopathy.

Fig 9: Swarm plot of SHAP values for the 9 most relevant variables and the sum of the rest for classification with a logistic regression model in patients with dysferlinopathies. The horizontal axis represents the SHAP values, which correspond to a measure of the contribution of each of the variables to the model decision for a category. A higher value corresponds to a positive impact on the prediction of the positive class and a negative value to a negative impact on the prediction of the positive class. Each dot corresponds to a patient in the test set. The colour of each dot corresponds to the value of that variable for that patient. Red indicates a higher degree of fatty infiltration and blue indicates a lower degree of fatty infiltration.