{"id":"8cdaf25f-10fc-473e-8fee-cf910ce0cfc3","arxiv_id":"2505.04073","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A PRISMA-based review categorizes 113 publications on medical natural language generation into text-to-text, image-to-text, and multimodal-to-text approaches and their clinical applications.","lead":"This preprint reviews 113 studies on natural language generation in healthcare and groups them by data modality, model architecture, clinical application, and evaluation method. A generalist reader would use it as a map of how generative AI is being applied to clinical documentation, summarization, dialogue, and synthetic data.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The modality taxonomy is internally inconsistent: structured EHR inputs are classified as 'text-to-text' despite the paper's own 'data-to-text' definition, so Table 1's prevalence counts do not reliably support the claimed multimodal map.","rationale":"The reader's weakest assumption identifies exactly the load-bearing weakness: the modality taxonomy is internally inconsistent. The paper needs the three categories to be exhaustive and mutually exclusive for Table 1's counts and the per-modality summaries to be interpretable. The Introduction's data-to-text versus text-to-text distinction makes structured EHR inputs non-linguistic, yet the later definition of text-to-text explicitly includes structured and semi-structured data. This is a direct, textually verifiable inconsistency, not a matter of external consensus. It affects the central quantitative artifact of the review: the modality prevalence counts in Table 1. If the structured-EHR studies are moved to a data-to-text category, the text-to-text denominator and the modality-specific architecture percentages change, and the claim of reviewing 'multiple modalities' loses its categorial foundation. The missing Supplementary Table S1 is a related but secondary problem: even if the taxonomy were fixed, the absence of per-study extraction data prevents independent audit. The firstness claim and the AlphaZero analogy are overstated but peripheral; they do not carry the review's evidentiary weight. Because the taxonomy and count inconsistencies are correctable with a documented reclassification and a released supplement, the appropriate verdict remains conditional rather than reject. I therefore agree with the reader and recommend no change to the verdict.","tokens_in":15546,"tokens_out":6044,"duration_ms":58361,"concrete_test":"Reconstruct the per-study modality labels for all 113 included studies using the Introduction's own definition (non-linguistic input counts as data-to-text, not text-to-text). Recode the structured-EHR studies — especially Lee et al. (ref 25), J Kurisinkel & Chen (ref 34), Soni & Dina (ref 92), and Guan et al. (ref 24) — and recompute Table 1. If the text-to-text denominator drops below 73, or if a fourth data-to-text category is needed to accommodate the recoded studies, then the published three-way taxonomy and the prevalence statistics do not support the claimed comprehensive modality-based review. Also independently sum each Table 1 modality column to verify the reported denominators.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Introduction distinguishes data-to-text (non-linguistic input) from text-to-text (text input), and lists structured EHR elements in database tables as a medical data modality. The 'Medical data modalities and text generation applications' section then says the 113 studies split into text-to-text, image-to-text, and multimodal-to-text, and defines text-to-text as including structured data (diagnosis codes, medications) and semi-structured tabular EHR entries. This directly contradicts the Introduction's data-to-text definition. The contradiction is not cosmetic: studies whose inputs are ICD codes, lab values, or structured EHR tables — e.g., Lee et al. (ref 25, FNN-to-LSTM from structured EHR codes), J Kurisinkel & Chen (ref 34, ICD codes to discharge instructions), and Soni & Dina (ref 92, structured/tabular EHR to progress notes) — are placed in the text-to-text rows of Table 1, even though the paper's own definition makes them data-to-text. Because the review's stated contribution is a modality-based map, and structured EHRs are explicitly named as one of the main medical data modalities, the three-way taxonomy omits the data-to-text category and instead dilutes 'text-to-text' with non-linguistic inputs. Consequently the Table 1 prevalence numbers (74 text-to-text studies, text says 73; 22 image-to-text; 17 multimodal-to-text) and the modality-specific qualitative summaries cannot cleanly support the central claim of a validated comprehensive overview. A separate counting inconsistency compounds this: Table 1's text-to-text rows sum to 74 while the text reports 73.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents a PRISMA-guided systematic review of natural language generation in healthcare. From 3,988 records screened through title/abstract and full-text stages, 113 studies are included and analyzed along four dimensions: input data modality, model architecture, clinical application, and evaluation metrics. The review reports that text-to-text, image-to-text, and multimodal-to-text are the three input-modality categories, identifies encoder-decoder and decoder-only transformers as dominant architectures, lists ROUGE/BLEU as the most common automatic metrics, and organizes applications into summarization, automated documentation, data augmentation, and medical dialogue. The central claim is that this is the first systematic review of NLG across multiple medical data modalities and applications.","tokens_in":15742,"tokens_out":6438,"duration_ms":57112,"significance":"The review is potentially useful as a structured inventory for researchers entering clinical NLG: the search protocol is described in unusual detail (seven databases, two-stage screening with two reviewers per article, training sessions, standardized extraction), and the paper ships concrete prevalence tables and a PRISMA flow diagram. If the modality taxonomy were coherent, Table 1 and Table 2 could serve as reference statistics. The claim to be the first multi-modality systematic review is plausible but not independently verifiable from the manuscript, and it is currently compromised by internal inconsistencies in the modality categories and in the counts, so the central contribution needs repair before the significance can be realized.","major_comments":[{"comment":"The modality taxonomy is internally inconsistent. The Introduction (p. 4) defines 'data-to-text' generation as generating text from non-linguistic input and explicitly lists structured EHR elements in database tables as a medical data modality; the Results section then groups the 113 studies into text-to-text, image-to-text, and multimodal-to-text only, and defines text-to-text as 'generating narrative text from textual input, including structured data (e.g., diagnosis codes, medications), semi-structured data (e.g., templates or tabular EHR entries)'. This classifies studies whose inputs are ICD codes, lab values, or structured EHR tables as text-to-text, even though the Introduction's own definition makes them data-to-text. Examples in the text are Lee et al. (ref 25, FNN-to-LSTM from structured EHR codes), J Kurisinkel & Chen (ref 34, ICD codes to discharge instructions), and Soni & Dina (ref 92, structured/tabular EHR to progress notes). Because the review's stated contribution is a modality-based map, and structured EHRs are named as one of the main medical data modalities, the three-way taxonomy omits the data-to-text category and the Table 1 prevalence counts do not cleanly support the claimed comprehensive overview. The authors should either reintroduce data-to-text as a separate category or explicitly redefine text-to-text to include non-linguistic structured inputs and justify that choice.","section":"Introduction and 'Medical data modalities and text generation applications'"},{"comment":"The prevalence numbers do not add up. The text states '61.64% (45/73) of studies used encoder-decoder transformer models, and 27.40% (20/73) of studies used decoder-only transformer models' in text-to-text generation, but Table 1 lists 74 text-to-text rows (1 GAN + 1 FNN/LSTM + 1 RNN/RNN + 3 GRU/GRU + 4 LSTM/LSTM + 1 LSTM/Transformer + 42 Transformer/Transformer + 1 no-encoder/LSTM + 20 decoder-only Transformer = 74), and no obvious reading yields 45 encoder-decoder transformer studies. The denominator should be a single consistent count, and all percentages should be recomputed from the corrected table.","section":"Table 1 and text after Table 1"},{"comment":"Table 2 is incomplete relative to the text. The text states that CIDEr is 'widely used in image-to-text generation evaluation' and that pairwise comparison is 'another widely used human evaluation metric', yet neither CIDEr nor pairwise comparison appears in Table 2's counts. Since evaluating metrics is one of the four stated review dimensions, Table 2 should either include these metrics with counts or the text should be revised to avoid claiming they are widely used.","section":"Table 2 and 'Image-to-Text Related Metrics'"},{"comment":"The three categories are not mutually exclusive under the given definitions. If structured EHR entries count as text, then a system taking both an image and structured clinical variables is simultaneously text-to-text and multimodal-to-text; if structured entries do not count as text, then the examples in the text-to-text paragraph contradict the definition. The taxonomy should specify whether modality labels are based on the input's linguistic status, its source, or the number of input channels, and Table 1's assignments should follow that rule.","section":"Definition of multimodal-to-text in Results"}],"minor_comments":[{"comment":"Reference list entries 79–81, 84–89, 92, 105, 106, 108, and 111 are incomplete (missing authors, venues, or titles); several appear as bare shared-task names. Please restore full bibliographic details.","section":"Reference list"},{"comment":"Figure 3's 'Text-to-Text Generation' panel includes a 'Structural EHR' table with PT_ID, TEST, ENC_DATE, VALUE; once the taxonomy is revised, the figure should be updated to match the corrected modality definitions.","section":"Fig. 3"},{"comment":"The sentence 'studies that about the potential bias, security, risks, interpretability, and AI ethics, but without concrete applications were not included' is ungrammatical and should be rewritten for clarity.","section":"Limitation section"},{"comment":"The sentence 'We have almost exhausted all the electronic data generated on this planet' is an unsupported overstatement and should be removed or supported with a citation.","section":"Discussion"},{"comment":"In Table 1, the row 'CNN +GAN Transformer' is ambiguous; clarify whether the encoder is CNN+GAN and the decoder is Transformer, and consider separating the entries.","section":"Table 1"},{"comment":"The sentence 'Both metrics have been utilized in our reviewed studies62–65' lists four references for two metrics; verify the citation grouping so each metric is supported by the intended studies.","section":"Evaluation metrics text"}],"recommendation":"major_revision","confidential_remarks":"The paper's usefulness depends on internal consistency of the modality taxonomy and on the accuracy of the prevalence counts. I see no evidence of misconduct or problematic citation patterns; the issues are fixable with reclassification, recomputation, and a supplemental mapping of each included study to a corrected modality category. The 'first systematic review' claim is difficult to verify from the manuscript alone, but I did not treat that as a disqualifying concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real contribution — the first systematic review I know of that covers medical NLG across text, image, and multimodal inputs, with a transparent PRISMA search and a useful breakdown of architectures, evaluation metrics, and applications. The prevalence numbers in Table 1 are not trustworthy as printed, but the problems are correctable and don't sink the review.\n\nWhat's genuinely good: the search protocol is described in enough detail (six databases plus ACL Anthology, 3,988 to 113, two-reviewer screening with third-party adjudication), and the paper gives a sensible qualitative map of how the field moved from RNN/GAN to transformer and LLM-based generation. The application sections (summarization, documentation, dialogue, augmentation) are organized and cite a broad set of studies. The limitation paragraph at the end is honest about excluding non-English work and non-application ethics studies.\n\nWhere the soft spots actually are. First, the taxonomy. The Introduction distinguishes data-to-text from text-to-text, and lists structured EHR elements as a non-linguistic modality. Then the results section defines text-to-text as including structured data (diagnosis codes, medications) and semi-structured tabular EHR entries, and places studies like Lee et al., J Kurisinkel & Chen, and Soni & Dina in the text-to-text rows. That is a direct internal contradiction, and it's load-bearing because the review's whole contribution is a modality-based map. The fix is straightforward: either add a data-to-text category, or argue explicitly why structured codes count as text. As it stands, Table 1's 74/22/17 split doesn't cleanly mean what the text says it means. Second, the count mismatch: Table 1 text-to-text rows sum to 74, but the text says 73 (61.64% of 73, etc.). Minor, but it erodes confidence in the extraction. Third, the supplement with per-study extraction (Supplementary Table S1) is not present in the preprint, so I can't verify the statistics. Fourth, the 'first study' claim and the AlphaZero analogy are overstatements; the firstness claim may be true but no search evidence for it is given, and AlphaZero is a poor analogy for synthetic clinical data.\n\nNet: the core review is solid and useful; the modality taxonomy and counting need a revision pass. I'd send it to peer review and require supplement plus a cleaned-up taxonomy. It belongs in the conversation once fixed.\n\nWho it's for: grad students and practitioners who want a fast map of medical NLG methods and benchmarks. I wouldn't cite the prevalence numbers until the taxonomy and counts are fixed.","headline":"A useful PRISMA-structured review of 113 medical NLG papers whose modality taxonomy has a genuine internal contradiction and some counting inconsistencies — fixable, not fatal.","tokens_in":16374,"tokens_out":2188,"would_cite":false,"duration_ms":22072,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This review claims the first systematic, multi-modality map of natural language generation in healthcare, covering 113 studies.","keywords":["natural language generation","healthcare","large language models","systematic review","clinical documentation","radiology report generation","medical dialogue systems","evaluation metrics"],"falsifier":"Take the 113 included studies and re-code each one by input type into four categories—free text, structured/tabular EHR data, images, and multimodal—then compare the resulting prevalence numbers with Table 1. If a substantial share of the text-to-text count moves into a data-to-text category (for example, studies that generate notes from diagnosis codes or lab values), the review's central taxonomic claim is contradicted by its own data.","tokens_in":15272,"feed_emoji":"🏥","tokens_out":4754,"duration_ms":39983,"temperature":0.7,"pith_summary":"This paper tries to establish a comprehensive, PRISMA-based map of natural language generation (NLG) in healthcare, claiming to be the first review to cover multiple medical data modalities and applications together. From 3,988 candidate articles, the authors screened down to 113 studies and grouped them by input type (text-to-text, image-to-text, multimodal-to-text), model architecture, clinical application, and evaluation method. The central result is a set of prevalence statistics: transformer encoder-decoder models dominate text-to-text generation, CNN+Transformer hybrids lead image-to-text work, and ROUGE, BLEU, and Likert-scale human ratings are the most common evaluation tools. A sympathetic reader would care because the paper turns a scattered field into reference numbers and a structured research agenda for clinical generative AI.","feed_headline":"First systematic map of medical text generation across 113 studies","feed_subtitle":"Review spans text, image, and multimodal inputs; transformers dominate and ROUGE/BLEU rule evaluation.","key_machinery":"The organizing machinery is a three-way taxonomy of input modality (text-to-text, image-to-text, multimodal-to-text) crossed with model architecture (encoder-decoder, decoder-only, and GAN-based), clinical application (summarization, documentation, dialogue, data augmentation), and evaluation method (n-gram, embedding-based, and human). This taxonomy is what turns 113 individual papers into aggregate prevalence statistics, so the whole review's descriptive claims rest on the categories being coherent and exhaustive.","core_discovery":"On its own terms, the paper's discovery is that NLG in healthcare has consolidated around transformer-based architectures across three input-modality classes, and that these classes map onto four clinical application families: summarization, automated documentation, dialogue generation, and data augmentation. The paper reports that 61.64% of text-to-text studies use encoder-decoder transformers, decoder-only transformers account for 27.40%, and image-to-text work relies heavily on CNN+Transformer hybrids, while evaluation is dominated by surface-level metrics (ROUGE in 74 studies, BLEU in 60) supplemented by Likert-scale human ratings in 36 studies. It further claims that this is the first systematic review of NLG that spans multiple medical data modalities and multiple healthcare applications, and it identifies synthetic clinical text generation as a promising route to privacy-preserving data scaling.","pith_inferences":["My inference: the text-to-text category probably absorbs data-to-text systems (structured EHR-to-narrative), because the paper's own introduction defines non-linguistic input as data-to-text; a cleaner four-way split would likely shift some prevalence numbers.","My inference: because most studies rely on n-gram overlap metrics, the review likely overstates how close the field is to clinically usable generation; a re-analysis correlating automatic scores with expert clinical judgment could test this.","My inference: the exclusion of non-English studies and of ethics/safety-only papers means the map is strongest for English-language, application-focused work; low-resource language NLG is a visible gap rather than a surveyed area.","My inference: the PRISMA-style pipeline could be rerun periodically to track how fast each modality and application category is growing, making this a baseline rather than a one-time snapshot."],"forward_implications":["If the taxonomy holds, Table 1's percentages become the field's reference statistics for where NLG research effort concentrates.","The dominance of ROUGE and BLEU suggests reported quality may miss clinical relevance and factual accuracy, motivating better evaluation metrics.","Synthetic clinical text generation, positioned as privacy-preserving data augmentation, is likely to grow as a way to scale training data without sharing patient records.","Multimodal-to-text generation, especially image-plus-structured-data fusion, is a rising pattern that could define the next phase of medical report generation.","The application map gives researchers and clinicians a structured menu for matching a clinical documentation problem to an existing architecture and evaluation setup."],"supporting_citations":[{"why":"Supplies the general definition of natural language generation that frames the review's scope.","marker":"[1]"},{"why":"Establishes the long-standing presence of NLG in health care, motivating the need for an updated review.","marker":"[2]"},{"why":"Provides an early GAN-based text-to-text method for generating synthetic electronic medical record text.","marker":"[24]"},{"why":"Supplies the early FNN-to-LSTM architecture that generates clinical note sections from structured EHR codes, anchoring the text-to-text category.","marker":"[25]"},{"why":"Offers an LSTM method for generating privacy-preserving synthetic clinical notes, grounding the data-augmentation discussion.","marker":"[33]"},{"why":"Represents the CNN-with-multi-attention plus LSTM approach to chest X-ray report generation, anchoring the image-to-text category.","marker":"[37]"},{"why":"Demonstrates the CNN-plus-memory-augmented-transformer architecture for clinical findings generation, illustrating the modern image-to-text shift.","marker":"[39]"},{"why":"Provides the knowledge-grounded multimodal transformer pipeline for radiology report generation, anchoring the multimodal-to-text category.","marker":"[42]"},{"why":"Supplies the MIMIC-CXR dataset that enables much of the chest X-ray report generation work reviewed.","marker":"[96]"},{"why":"Supplies the IU X-Ray dataset used for training and evaluating radiology report generation models.","marker":"[97]"}],"fun_headline_variants":["113 studies map medical text generation","First systematic NLG review: text, images, multimodal","Transformers dominate medical NLG, ROUGE rules","Medical NLG: four uses, three input types, one review","Synthetic data via NLG: privacy-preserving scaling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole classification depends on the three input-modality categories being exhaustive and on structured EHR data counting as text-to-text input, even though the introduction defines generation from non-linguistic input as data-to-text; if that labeling is inconsistent, the prevalence table and the grouped summaries stop being a reliable map of the field.","fun_headline_variants_meta":{"raw":{"variants":["113 studies map medical text generation","First systematic NLG review: text, images, multimodal","Transformers dominate medical NLG, ROUGE rules","Medical NLG: four uses, three input types, one review","Synthetic data via NLG: privacy-preserving scaling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000111,"raw_usage":{"total_tokens":1040,"prompt_tokens":913,"completion_tokens":127,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":49}},"tokens_in":529,"tokens_out":127,"duration_ms":2061,"temperature":1.0,"reasoning_tokens":49,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:37:59.789846+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the 113 included studies and re-code each one by input type into four categories—free text, structured/tabular EHR data, images, and multimodal—then compare the resulting prevalence numbers with Table 1. If a substantial share of the text-to-text count moves into a data-to-text category (for example, studies that generate notes from diagnosis codes or lab values), the review's central taxonomic claim is contradicted by its own data.","supporting_citations":[],"review_version":1}