“…Data quality issues and potential security risks have been continuously emphasized in the KG application [ 95–97 , 99 , 100 ]. The multifaceted issues of data quality in data extraction and curation, bias, data poisoning and dataset update have been considered substantially by us and others and mitigated by diverse solutions, including the new NLP technology development [ 101 , 102 ], domain experts assignment [ 103 ], stepwise bias-mitigating framework [ 104 ], adversarial training [ 105 ] and automatic updating system for primary dataset sources, respectively. The ontology problems, such as acronyms, homonyms and the hierarchy of biomedical terminology, have been improved, but we still encounter problems in KG-based models, requiring a unified multimodal biomedical ontology system for ML [ 101 , 106 , 107 ].…”