Fig 1.
Existing CERT, CLEAR, and PACE algorithms are applied to the Ajou University Hospital EHR dataset to extract features for ML algorithms for combining ADR signals. EU-SPC and SIDER databases are used to define the ADR reference dataset.
Table 1.
Features derived from three ADR signal detection algorithms.
Table 2.
Performance of ML models and previous ADR signal detection methods.
Table 3.
Summary table of Tukey’s HSD post-hoc test results among ML models.
Fig 2.
AUROCs of ML algorithms and original ADR signal detection algorithms.
The best AUROCs (A) and average AUROCs (B) of each algorithm are shown among 10 experiments with tenfold cross-validation. No significant difference is observed in the ML models; however, the AUROCs of the ML models are much larger than those of the original methods.
Fig 3.
Important features in random forest algorithm.
The importance of features is expressed in terms of Gini importance by color during 10 experiments with tenfold cross-validation. The blue color implies more importance and the yellow color, less importance. The top 10 important features are marked by red boxes.
Fig 4.
Coefficients calculated in logistic regression.
The coefficients of features are expressed in color during 10 experiments with tenfold cross-validation. Red color indicates positive coefficients and blue color, negative coefficients.