Table 2.

Categorization of papers reviewed regarding data sources and machine learning methods used for four causal inference paradigms

Machine learning methods		Data source
Causality assessment scales
SVM, Bayesian classifier, decision tree and/or Random Forest [86–88]		SRS [86, 87] Social media [88] RWD [89–91]
Regression [89]
Neural network [90, 91]
Propensity score matching
SVM, Bayesian classifier, decision tree and/or Random Forest [92, 93]		RWD [20, 93–99] Simulated data [92, 100]
Ensemble (boosting/bagging) [20, 92, 94, 95, 99]
Regression [20, 94, 95, 97]
Neural network [96, 98, 100]
Graph-based causal inference
SVM, Bayesian classifier, decision tree and/or Random Forest
Link prediction [101]		Knowledge bases [13, 102–104] RWD [101, 105–112]
Recommendation systems [109]
Classification [110]
Graph embedding
Link prediction [13, 102]
Recommendation Systems [105]
Regression
Classification [103]
Neural network
Link prediction [104]
Recommendation systems [106]
Predictive modeling [107, 108]
Ensemble (boosting/bagging)
Classification [111]
Link prediction [112]
Instrumental variables
Clustering [113]		RWD [113–116] Simulated data [16, 114, 115, 117–119] Social network data [115]
Decision tree [16]
Neural network [114–117]
New algorithms [118, 119]

RWD real-world data, SRS spontaneous reporting system, SVM support vector machine

Papers for “propensity score matching” and “instrumental variables” are not applied in the field of pharmacovigilance. Papers for “graph-based causal inference” still lacks a clear causal interpretation from a graph perspective