Table 4.

Summary of models for question answering (QA) over electronic health records.

Papers	Model
Pampari et al [26]	For QA task: DrQA’s document reader and a multiclass logistic regression model for predicting class. For question-to-logical form task: a sequence-to-sequence model is used with attention paradigm
Moon et al [27]	Clinical BERTa model with incremental masking
Oliveira et al [38]	BioBERTpt
Yue et al [42]	For QA task: DrQA’s DocReader and ClinicalBERT For question generation task: QPPb module is used with base question generation models (NQGc, NQG++, and BERT-SQGd)
Hamidi and Roberts [48]	ChatGPT (versions 3.5 and 4), Google Bard, and Claude
Fleming et al [49]	6 language models: GPT-4 (32 K tokens+multistep refinement), GPT-4 (32-K tokens), GPT-4 (2K tokens), Vicuña-13B (2K tokens), Vicuña-7B (2K tokens), and Vicuña-7B (2K tokens)
Mahbub et al [50]	Baseline models: 4 state-of-the-art pretrained language models—BERT, BioBERT, BlueBERT, and ClinicalBERT for QA. Modeling with transfer learning: sequential learning and adversarial learning
Dada et al [51]	G-BERT and GM-BERT
Roberts and Patra [57]	Hybrid semantic parsing method, uses rule-based methods along with a machine learning–based classifier.
Rawat and Li [59]	Uses multilevel attention layers along with local and global context while answering questions
Rawat et al [60]	Multitask learning with BERT and ERNIE [76] as the base model
Wen et al [64]	BERT model trained on different data sources
Soni and Roberts [65]	BERT, BioBERT, clinical BERT, and XLNet
Mairittha et al [66]	BERT (large, uncased, whole word masking), BERT fine-tuned on SQuADe benchmark, BioBERT, and an extended BioBERT fine-tuned on unstructured EHR data
Moon et al [67]	ClinicalBERT model fine-tuned on SQuAD-why dataset
Li et al [68]	Clinical-Longformer and Clinical-BigBird language model
Yang et al [69]	GatorTron language model
Lehman et al [73]	12 different language models (T5-Base, Clinical-T5-Base-Ckpt, Clinical-T5-Base, RoBERTa-Large, BioClinRoBERTa, GatorTron, T5-Large, Clinical-T5-Large, PubMedGPT, T5-XL, Flan-T5-XXL, and GPT-3)
Kang et al [70]	KALAf
Wang et al [5]	TREQSg
Raghavan et al [8]	Min et al [77] for sequence-to-sequence task along with ParaGen and ParaDetect model
Pan et al [62]	Medical text-to-SQL model
Soni and Roberts [63]	Tranx, Coarse2Fine, transformer, and lexicon-based
Tarbell et al [71]	T5 language model for question-to-SQL task, along with data augmentation method for back-translation
quEHRy [72]	End-to-end EHR QA pipeline with concept normalization (MetaMap), time frame classification, semantic parsing, visualization with question understanding, and query module for FHIRh mapping and processing
Kim et al [39]	Program-based model
Wang et al [40]	Attention-based aspect reasoning
Park et al [41]	Seq2Seq model [78] and TREQS [5]
Schwertner et al [58]	ENSEPROi framework
Bae et al [61]	Unified encoder-decoder architecture that uses input masking
Bardhan et al [25]	MultimodalEHRQA

^aBERT: Bidirectional Encoder Representations from Transformers.

^bQPP: question phrase prediction.

^cNQG: Neural Question Generation.

^dBERT-SQG: BERT-Sequential Question Generation.

^eSQuAD: Stanford QA dataset.

^fKALA: Knowledge-Augmented Language model Adaptation.

^gTREQS: Translate-Edit Model for Question-to-SQL.

^hFHIR: Fast Healthcare Interoperability Resources.

ⁱENSEPRO: Ensino de Serviços Proativos (in Portuguese), which translates to Teaching Proactive Services.