3.3.3 Model interpretation
Construct the S HAP value variable importance map (Figure 9A) of the prediction model, which includes a total of 5 features, sorted according to their influence on the level of TNF-α, the higher the absolute value of SHAP value, the greater the effect on TNF-α powerful. The results showed that the characteristics of ”Type of disease”, ”Animal species” and ”Dose” had a higher contribution to the improvement of TNF-α level (Figure 10). In the constructed SHAP value summary graph, the higher the SHAP value corresponding to a feature, the greater the risk of increasing TNF-α improvement level, and vice versa. In addition, the color of the point represents the improvement rate of TNF-α, and the darker the color, the better the effect. The results showed that, among the categorical variables, for respiratory, urinary, neurological and digestive diseases included in the ”type of disease” feature, icariin and its derivatives had a more positive effect on the TNF-α efficacy of these systemic diseases compared with the baseline values. Meanwhile, ”Wister mice” and ”BALB/c mice”, as species in the ”Animal species” characterization, contributed more to the TNF-α efficacy compared with the baseline values. And ”female” as a variable in the ”sex” characteristic was also beneficial in improving the efficacy. (Figure 10B). As for continuous variables, when the ”course” was greater than 31.22 days or the ”dose” was greater than 27.52 mg/kg/day, icariin and its derivatives contributed more positively to the rate of improvement in TNF-alpha compared to baseline values (Figure 10C).