{"id":"ad15032c-c7f8-4c15-b16e-b509e212a35d","arxiv_id":"2506.14789","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"AZT1D is a publicly released real-world dataset with CGM, insulin, carbohydrate, and device-mode data from 25 T1D patients on automated insulin delivery, including rare bolus-level detail.","lead":"A new dataset, AZT1D, captures continuous glucose, insulin, carbohydrates, and pump mode from 25 people with type 1 diabetes using automated insulin delivery systems. It stands out for its detailed bolus-level insulin records, which could help train AI models for personalized diabetes care.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated OCR extraction of basal rates and device modes is the key unsecured link; without an audit or error rate, the derived columns of AZT1D may not be trustworthy.","rationale":"I agree with the reader that the unvalidated OCR/PDF extraction pipeline is the most load-bearing weakness. The dataset's novelty rests on combining CGM, insulin events, carbohydrates, and device mode; device mode and basal rate are derived from PDFs with no reported validation, and the graphical nature of the source makes OCR particularly error-prone. The internal inconsistency between the stated average duration and the total hours/CGM entries is also real and should be corrected, but it is more likely a reporting typo than a fundamental flaw, whereas a systematic OCR error would directly corrupt the derived columns. The paper does provide positive evidence: a public DOI, a concrete patient table, and two application papers using the dataset, which support the existence and utility of the resource. However, because the derived fields are not validated and the described duration is internally inconsistent, the current conditional verdict is appropriate. No change to the reader's verdict is needed; the same concern justifies keeping the paper conditional pending a public validation report.","tokens_in":4484,"tokens_out":4440,"duration_ms":45458,"concrete_test":"Obtain the original Tandem PDF exports for a subset of patients, for example 5 of the 25, and manually annotate device mode and hourly basal rate for each 24-hour period. Compare these manual annotations against the released AZT1D columns after applying the same timestamp alignment, and report per-field agreement and timestamp misalignment rates. If device-mode agreement is below 95% or any basal rate differs by more than 0.05 U/h, the OCR pipeline is not validated and the derived fields should be regenerated or marked provisional. In the same pass, recompute total monitoring hours from patient-level durations to resolve the 26-day versus 26,707-hour inconsistency.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central value of AZT1D depends in part on the derived basal-rate and device-mode columns, which are advertised as dataset components and feed downstream ML analyses. Section III-A states that these fields are extracted from clinic PDFs by cropping regions and applying OCR, but no manual audit, OCR error rate, or comparison against original pump exports is reported. The OCR step is especially fragile for device mode: in Tandem t:slim X2 reports this information is rendered graphically as shaded intervals, so character recognition alone cannot reliably assign regular/sleep/exercise labels. Basal rates are hourly and then repeated across five-minute CGM intervals, so a single OCR error propagates to 12 rows and persists for up to an hour. Timestamp unification between CSV timestamps and PDF time axes is also unvalidated. If these columns contain systematic errors, the dataset is not a clean structured real-world resource and the claimed advantage over existing datasets is weakened. This concern is distinct from the bolus CSV fields, but device mode and basal rate are part of the advertised contribution. An additional red flag is the internal inconsistency between the stated average duration of 26 days and the total of 26,707 hours / 320,488 CGM entries, which imply about 44.5 days per patient; this should also be corrected while auditing the pipeline.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"AZT1D is presented as a publicly available real-world dataset from 25 individuals with type 1 diabetes using automated insulin delivery (Tandem t:slim X2 with Control IQ and Dexcom G6 Pro CGM). The dataset contains CGM readings, bolus insulin logs with detailed type and correction-specific amounts, carbohydrate entries, basal rates, and device mode classifications. The paper describes the study design, the OCR-based preprocessing pipeline for extracting basal rates and device modes from clinic PDFs, the dataset schema, and a basic statistical analysis of glycemic events in elderly patients. Two prior studies by the authors are cited as use cases of the dataset. The abstract and introduction claim that this is the only publicly available dataset with this level of bolus-event detail.","tokens_in":4759,"tokens_out":3349,"duration_ms":33508,"significance":"If the dataset is properly validated, it is a useful contribution to the T1D machine-learning community. The granular bolus fields (total dose, bolus type, correction-specific amounts) are indeed rare in public datasets, and the multi-week real-world collection with device-mode context supports a range of ML tasks. The authors also share demographic and clinical data and state an intention to release daily and monthly visualizations, which aids reproducibility. The paper's main value, however, depends on the reliability of the derived basal-rate and device-mode columns and on the accuracy of the reported dataset statistics. Those points are not yet established, so the contribution is currently conditional on additional validation.","major_comments":[{"comment":"The reported duration is internally inconsistent. The abstract says data were obtained over 6 to 8 weeks per patient, Section II says participants contributed data over an average duration of 26 days, and the reported 320,488 CGM entries at 5-minute intervals total 26,707 hours, which implies about 44.5 days per patient on average. Please reconcile these numbers, clarify what the 26-day figure refers to, and report the exact per-patient minimum, maximum, and median observation windows. The multi-week coverage is a central selling point and must be accurately described.","section":"Section II / Abstract"},{"comment":"The OCR-based extraction of basal rates and device modes from clinic PDFs is not validated. The paper states that these fields are obtained by cropping informative regions and applying OCR, but it provides no manual audit, OCR error rate, or comparison against original pump exports. This is load-bearing because device mode in Tandem t:slim X2 reports is rendered graphically as shaded intervals, so character recognition alone may not reliably assign regular/sleep/exercise labels, and an error in an hourly basal value propagates to 12 five-minute records. Please add a validation subsection that reports error rates, describes how the graphical device-mode regions were segmented and verified, and shows agreement against a ground-truth subset.","section":"Section III-A"},{"comment":"The statistical analysis reports durations and percentages of hyperglycemia and hypoglycemia without defining the glucose thresholds used to classify these events. For example, the statement that hyperglycemia accounted for 28.6% of evening time is not interpretable unless the reader knows whether the threshold is, say, >180 mg/dL or another value. Please state explicit thresholds for hypo- and hyperglycemia (e.g., <70 mg/dL and >180 mg/dL, following standard CGM consensus) and cite the relevant clinical definitions. The same definitions must be applied consistently in Figure 3 and in any released per-patient summaries.","section":"Section IV"}],"minor_comments":[{"comment":"There is a missing space in 'we presentAZT1D' in the abstract.","section":"Abstract"},{"comment":"The phrase 'O pticalC haracterR ecognition' has unintended spacing; please write 'OCR' or 'Optical Character Recognition' with normal spacing.","section":"Section III-A"},{"comment":"The demographic table lacks a caption and a definition of the A1c column; please add a caption and state that A1c is measured in percentage (DCCT units).","section":"Table I"},{"comment":"The 'monthly report' mentioned in the text is not shown or referenced to a repository location; please indicate where these materials are available or remove the claim.","section":"Section IV"},{"comment":"Figure 3 has no axis labels; please add a y-axis label such as 'Total duration (minutes)' and clarify in the caption that 'elderly' is defined as age ≥65 years and how many patients are included.","section":"Figure 3"},{"comment":"References [7] and [8] are prior works by the same group that used this dataset; please state explicitly that these are applications of AZT1D rather than independent validations, so readers can calibrate the reported 76.6%/86% and RMSE/MAE improvements.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"The dataset description is a valuable resource, but the OCR validation gap and the duration inconsistency are load-bearing and require correction in the manuscript. The fixes are within the scope of a revision: a validation subsection and a precise data-duration table would address the main concerns. The paper is not ready for acceptance in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick read on AZT1D: the dataset fills a real gap. Existing public T1D datasets like OhioT1DM and Hupa-UCM don't give you bolus type, correction-specific insulin amounts, and device mode at this granularity. If the data is clean, it enables granular glucose prediction and treatment policy work that wasn't straightforward before. The paper also does a few things right: it documents the fields, reports demographics, gives IRB details, and actually ships the data via Mendeley. That's more than many dataset papers do.\n\nThe problems are concentrated in the derived columns. Section III-A says basal rates and device modes are extracted from clinic PDFs by cropping and OCR. There's no error rate, no manual audit, no comparison against original pump exports. That's a real gap, because device mode in Tandem t:slim X2 reports is rendered as shaded intervals; OCR alone probably can't classify regular/sleep/exercise reliably. And a single basal-rate OCR error propagates to 12 five-minute rows. Timestamp unification between the CSV and PDF time axes is also unvalidated. If these columns are systematically wrong, the dataset's advertised advantage over existing resources shrinks.\n\nThe internal numbers also don't line up. The abstract says 6–8 weeks per patient; Section II says an average of 26 days; but 320,488 CGM entries at 5-minute intervals across 25 patients implies about 44.5 days per patient. That's not a rounding difference. The authors need to reconcile this and explain what the actual monitoring window is.\n\nSmaller issues: the statistical analysis uses hyper/hypoglycemia thresholds without defining them. The cohort is small (25) and skewed to elderly patients, which they acknowledge by focusing the analysis there, but it limits generalization. The 'Use Cases' section cites two earlier papers by the same authors using this dataset; that's fine, not circular, but it would help to state the dataset has already been used in published work.\n\nNone of this disproves the dataset's existence. The bolus CSV fields likely come straight from the pump and are probably sound. The OCR-derived fields are the load-bearing risk. With an audit, a measured error rate, and fixed inconsistencies, this could be a solid resource. As it stands, I'd want that validation before building anything on it.\n\nRecommendation: send it to peer review, but with a clear request for the OCR validation and duration reconciliation. A serious referee can sort this out.","headline":"AZT1D's bolus-level detail is genuinely useful, but the unvalidated OCR pipeline and internally inconsistent duration figures need to be fixed before the derived columns can be trusted.","tokens_in":5272,"tokens_out":2469,"would_cite":false,"duration_ms":22130,"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":"The paper introduces AZT1D, a public real-world dataset of 25 people with type 1 diabetes using automated insulin delivery, with continuous glucose, insulin, carb, and device-mode data, and claims it is the only public dataset with such…","keywords":["type 1 diabetes","automated insulin delivery","continuous glucose monitoring","bolus insulin","real-world dataset","machine learning","optical character recognition","device mode"],"falsifier":"Re-extract the hourly basal rates and device modes from a random sample of the original clinic PDF reports using independent manual transcription or a second OCR tool, then compare against the published dataset; a mismatch rate above a few percent would show the dataset is not as cleanly structured as claimed.","tokens_in":4334,"feed_emoji":"💉","tokens_out":5728,"duration_ms":50950,"temperature":0.7,"pith_summary":"The paper introduces AZT1D, a publicly released dataset of 25 people with type 1 diabetes who use automated insulin delivery systems. It unifies continuous glucose readings, insulin administration logs, carbohydrate entries, and device modes into a single timestamped timeseries spanning several weeks per patient. The authors' central claim is that AZT1D is the only public dataset with granular bolus insulin detail—total dose, bolus type, and the correction-specific portion of each event. That matters because such linked, real-world data is what machine-learning models need for personalized dosing advice and glucose prediction.","feed_headline":"25 T1D patients' real-world glucose and insulin data go public","feed_subtitle":"Granular bolus-event fields make it the only public set of its kind, the authors argue.","key_machinery":"The central object is the AZT1D dataset itself, a structured timeseries in which each row carries an event timestamp, device mode, basal insulin rate, bolus type, correction-delivered portion, total bolus insulin, food-delivered portion, carbohydrate size, and CGM glucose value. Two mechanisms build it: timestamp alignment that merges the pump's five-minute glucose readings with the pump's event logs and carb entries, and an OCR-with-coordinate-cropping pipeline that extracts hourly basal rates and device modes from clinic PDF reports. The granular bolus fields are what the paper argues no other public dataset provides.","core_discovery":"On its own terms, the contribution is a new resource rather than a new finding about diabetes. AZT1D contains 320,488 continuous glucose monitoring entries collected over a combined 26,707 hours, alongside insulin delivery events, carbohydrate sizes, and a device-mode label for each record. The paper's distinctive assertion is that this is the only publicly available dataset with this level of detail on bolus insulin events, separating standard, correction, and automatic boluses and reporting the correction and food portions of each dose. To create the dataset, the authors aligned timestamps across data sources and used optical character recognition on clinic-generated PDF reports to recover hourly basal rates and device modes. Their aggregate analysis of patients 65 and older shows hyperglycemia rising through the day to peak in the evening and hypoglycemia most frequent in the early morning.","pith_inferences":["Because the two cited validation studies were run on this same dataset, their reported gains (76.6 percent valid counterfactual explanations and a 23 percent RMSE reduction in glucose prediction) inherit any errors in the OCR extraction and timestamp alignment.","The uniqueness claim is time-limited: new public datasets with equally detailed bolus fields would undo it, so the claim should be understood as a statement about the current landscape.","A directly testable extension would be an ablation study that trains glucose-prediction models with and without the device-mode and correction-delivered fields to quantify how much these novel fields actually improve accuracy.","If the OCR pipeline ever fails on a subgroup of patients, analyses restricted to the CSV-derived fields (CGM, bolus logs, carb sizes) would remain reliable, making it worth reporting which fields depend on the PDF extraction."],"forward_implications":["Machine-learning models for glucose prediction and insulin dosing can now be trained on naturalistic multi-week data instead of simulated or small-cohort datasets.","The correction/food split of bolus events enables counterfactual recommendation systems that tell patients how to change dosing or meals to avoid hyperglycemia.","The temporal pattern seen in elderly patients—evening hyperglycemia and early-morning hypoglycemia—points toward time-specific monitoring and management strategies.","The dataset's temporal structure is suitable for offline reinforcement learning to learn personalized insulin policies from historical records.","Expanding the cohort to about 100 patients is planned, which would strengthen the generalizability of conclusions drawn from AZT1D."],"supporting_citations":[{"why":"Provides the prior public T1D dataset that AZT1D compares against; the scarcity claim depends on this dataset lacking the same bolus detail.","marker":"[5]"},{"why":"Another existing public T1D dataset representing the prior state of the art; used as evidence that detailed bolus events are missing.","marker":"[6]"},{"why":"A counterfactual-explanation study run on AZT1D, used to demonstrate the dataset supports personalized recommendation use cases.","marker":"[7]"},{"why":"A glucose-prediction model with a custom loss function validated on AZT1D, used to show the dataset's utility for predictive modeling.","marker":"[8]"}],"fun_headline_variants":["New dataset: 25 patients, 320K glucose readings, detailed bolus events","First public T1D dataset with granular bolus insulin details","AZT1D: Real-world T1D data with rare bolus-level detail","25 T1D patients, 6-8 weeks, granular insulin data now public","Dataset fills gap: detailed bolus events in type 1 diabetes care"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The dataset's basal rates and device-mode labels come from an OCR and cropping pipeline on clinic PDF reports, so any systematic OCR error or timestamp misalignment would propagate into every field derived from those PDFs.","fun_headline_variants_meta":{"raw":{"variants":["New dataset: 25 patients, 320K glucose readings, detailed bolus events","First public T1D dataset with granular bolus insulin details","AZT1D: Real-world T1D data with rare bolus-level detail","25 T1D patients, 6-8 weeks, granular insulin data now public","Dataset fills gap: detailed bolus events in type 1 diabetes care"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000626,"raw_usage":{"total_tokens":2871,"prompt_tokens":897,"completion_tokens":1974,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":1871}},"tokens_in":513,"tokens_out":1974,"duration_ms":12719,"temperature":1.0,"reasoning_tokens":1871,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:22:08.088354+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-extract the hourly basal rates and device modes from a random sample of the original clinic PDF reports using independent manual transcription or a second OCR tool, then compare against the published dataset; a mismatch rate above a few percent would show the dataset is not as cleanly structured as claimed.","supporting_citations":[{"cited_title":"Limbert, A","cited_arxiv_id":null,"evidence_quote":"Provides the prior public T1D dataset that AZT1D compares against; the scarcity claim depends on this dataset lacking the same bolus detail."},{"cited_title":"Marling and R","cited_arxiv_id":null,"evidence_quote":"Another existing public T1D dataset representing the prior state of the art; used as evidence that detailed bolus events are missing."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A counterfactual-explanation study run on AZT1D, used to demonstrate the dataset supports personalized recommendation use cases."},{"cited_title":"Arefeen, S","cited_arxiv_id":null,"evidence_quote":"A glucose-prediction model with a custom loss function validated on AZT1D, used to show the dataset's utility for predictive modeling."}],"review_version":1}