AI-READI Flagship Dataset

• AI-READI Flagship Dataset is a large, multimodal resource curated to advance Type 2 Diabetes research by aggregating biosensor and clinical signals from over 1,000 diverse participants.
• It supports various task-specific views such as blood glucose forecasting, binary stress prediction, and privacy-preserving synthetic data generation.
• The dataset drives precision health research by integrating heterogeneous sensor data, employing robust preprocessing, and benchmarking continual learning methods.

Searching arXiv for papers on the AI-READI Flagship Dataset and closely related uses. I’m checking arXiv entries relevant to AI-READI and its downstream modeling papers. The AI-READI Flagship Dataset is a large, publicly available, multimodal resource curated to advance AI and machine learning research on Type 2 Diabetes Mellitus (T2DM). In current arXiv work, it supports heterogeneous analytical regimes, including blood glucose forecasting from continuous glucose monitoring and activity signals, binary stress prediction from daily or nightly wearable summaries, LLM-prompted counterfactual generation, and privacy-preserving synthetic data generation from person-day tabular abstractions. Across these uses, the dataset functions less as a single fixed benchmark matrix than as a broad multimodal substrate from which task-specific views are constructed (Farahmand et al., 14 Feb 2025, Soumma et al., 7 Jul 2025, Espinosa-Dice et al., 24 Dec 2025).

1. Purpose, scope, and cohort basis

The Flagship AI-READI dataset was curated “to advance AI and machine learning research on Type 2 Diabetes Mellitus” and includes individuals with and without T2DM. It is described as balanced across sex, race, and diabetes severity, although the studies summarized here do not report more granular demographic distributions beyond that characterization. The full dataset was collected from 1,067 participants across three U.S. sites, spanning the spectrum from healthy individuals to people with T2DM (Farahmand et al., 14 Feb 2025).

Within published downstream uses, the dataset is explicitly framed as an NIH-funded initiative that aggregates rich biosensor and clinical signals to enable precision health modeling. One forecasting study retains 896 participants after filtering for missing data, partitioned into healthy, prediabetes, T2DM on oral medication, and T2DM on insulin cohorts. Another study recasts the resource into 4,537 person-days with 108 features for synthetic data generation. These distinct materializations indicate that AI-READI is both cohort-rich and representation-flexible, with the same underlying program supporting raw or semi-raw temporal modeling, daily summary modeling, and structured tabular synthesis (Farahmand et al., 14 Feb 2025, Espinosa-Dice et al., 24 Dec 2025).

2. Modalities and observational structure

The dataset includes multiple sensing and clinical sources. In the blood glucose forecasting formulation, the principal devices are the Dexcom G6 for continuous glucose monitoring and the Garmin Vivosmart 5 for physical activity and heart rate variability, with a stress index on a $0\text{--}100$ scale computed by the wearable from HRV. The broader dataset also includes environmental sensing from the LeeLab Anura sensor, as well as survey data, clinical assessments, and retinal imaging, although those additional sources are not used in that particular forecasting model (Farahmand et al., 14 Feb 2025).

In the stress-prediction and counterfactual setting, the paper emphasizes twelve tabular features derived from Garmin Vivosmart 5 and Dexcom G6 summaries, including deep sleep percentage, REM sleep percentage, blood glucose in mg/dL, time-in-range, number of hyperglycemia events, average daily step count, and a device-derived stress score. In the synthetic-data setting, AI-READI is further expanded to include wearable summaries, clinical and ECG features, and diagnosis indicators at the person-day level (Soumma et al., 7 Jul 2025, Espinosa-Dice et al., 24 Dec 2025).

Source	Reported content	Use in current studies
Dexcom G6 CGM	Real-time blood glucose; every 5 minutes	Forecast target; daily glucose summaries
Garmin Vivosmart 5	Activity, sleep, HRV-derived stress, steps	Forecast covariates; stress features
LeeLab Anura sensor	Air quality, temperature, environmental signals	Present in dataset; not used in AttenGluco
Clinical, survey, ECG, retinal imaging	Clinical assessments, diagnoses, eye imaging	Used selectively in tabular studies

For the forecasting study, monitoring duration is ten days per participant, and the CGM sampling rate is every 5 minutes. Physical activity is represented through daily step counts from an accelerometer, with walking events treated as irregular and supplemented by constructed walking time intervals defined as durations between consecutive walking events. Heart rate, stress, and environmental sampling rates are not explicitly reported in that paper. This asymmetry in reporting is characteristic of current AI-READI use: some modalities are described at the raw stream level, while others appear only after aggregation into task-specific features (Farahmand et al., 14 Feb 2025).

3. Derived representations and preprocessing conventions

A central characteristic of AI-READI in the present literature is that the dataset is repeatedly re-expressed into task-specific analytical views. In AttenGluco, the operative inputs are $x_g$ for CGM, $x_{ws}$ for walking steps, and $x_{wi}$ for walking intervals, with the prediction target given by CGM-derived blood glucose level. Input windows are aligned to the CGM timeline because the output is defined on CGM values. Missing values are handled via interpolation, the preprocessing includes “normalization for consistency,” and fusion across mismatched sampling rates is performed with cross-attention rather than hard resampling. The precise interpolation method, timestamp alignment details beyond CGM anchoring, and normalization scheme are unspecified (Farahmand et al., 14 Feb 2025).

In SenseCF, the same broader data ecosystem is transformed into structured feature vectors $x \in \mathbb{R}^d$ with $d = 12$. The representation is tabular rather than sequential, and features are daily or nightly summaries rather than sliding windows. Age, sex, and medication status are treated as immutable variables; sleep, glucose, stress score, and activity features are mutable. The paper does not describe specific filtering, resampling, or normalization for training, although its evaluation uses $L_2$-normalized continuous features and Hamming distance for categorical features (Soumma et al., 7 Jul 2025).

In RLSyn, AI-READI v2.0 is transformed into a person-day table with 4,537 rows and 108 features. The reported breakdown is 16 wearable-derived features, 56 continuous clinical or ECG measurements, and 36 diagnosis variables. Time-series data from multiple sensors are temporally aligned to a common grid; short gaps are interpolated; and person-days with insufficient sensor coverage or extended non-interpolable gaps are excluded. For privacy analysis, features are normalized to $[0,1]$ before Euclidean distance computations. This establishes a third, higher-dimensional view of AI-READI oriented toward structured synthetic data generation rather than direct physiological forecasting (Espinosa-Dice et al., 24 Dec 2025).

4. Benchmark task formulations

The most fully specified temporal benchmark on AI-READI is blood glucose forecasting. Both AttenGluco and its multimodal 1D-CNN + LSTM baseline use a sliding historical window of 400 minutes, or 6.66 hours, and predict at horizons of 5, 30, and 60 minutes. Because CGM is recorded every 5 minutes, these correspond to $m = 1$, $6$, and $x_g$0 forecasted CGM samples. The task is formalized as

$x_g$1

with $x_g$2 denoting the set of sensor streams over sampling duration $x_g$3. Training uses 300 epochs, Adam with learning rate $x_g$4, and Mean Squared Error as the objective, with five runs per model to report average performance across subjects (Farahmand et al., 14 Feb 2025).

The stress-prediction formulation is different in both label semantics and temporal abstraction. SenseCF defines a binary outcome: high stress $x_g$5 versus moderate stress $x_g$6. Ground-truth labeling originates from Garmin-derived stress features, but the thresholds converting device stress scores into the binary label are not specified. Counterfactual generation is then posed as finding a minimally altered, plausible feature vector $x_g$7 such that $x_g$8, subject to immutable features remaining fixed and mutable features staying within training-data ranges. The paper reports validity, sparsity, plausibility, and a mixed distance metric, with plausibility defined over whether generated feature values remain within empirical training-set ranges (Soumma et al., 7 Jul 2025).

The synthetic-data benchmark in RLSyn evaluates AI-READI under privacy, utility, and fidelity criteria. Utility is measured by synthetic-to-real type 2 diabetes classification, where a model is trained on synthetic data and evaluated on a real test set. Fidelity includes real-to-synthetic AUC, latent-space similarity via Normalized Mutual Information after PCA and $x_g$9-means, column-wise correlation differences computed from Pearson correlation matrices, and a dimension-wise difference built from absolute prevalence differences for categorical variables and normalized Wasserstein distances for continuous ones. Privacy is evaluated with a membership inference attack based on nearest-neighbor Euclidean distances after $x_{ws}$0 normalization (Espinosa-Dice et al., 24 Dec 2025).

A representative counterfactual example in the stress study illustrates the semantics of these tabular features. A factual instance corresponds to an 81-year-old labeled “stressed” with deep sleep $x_{ws}$1, REM sleep $x_{ws}$2, blood glucose $x_{ws}$3 mg/dL, average steps $x_{ws}$4, stress level $x_{ws}$5, time-in-range $x_{ws}$6, and one hyperglycemia event. The LLM-suggested counterfactual preserves immutable variables and changes deep sleep to $x_{ws}$7, REM sleep to $x_{ws}$8, and blood glucose to $x_{ws}$9 mg/dL, thereby flipping the predicted label to moderate stress (Soumma et al., 7 Jul 2025).

5. Reported empirical behavior across tasks

For blood glucose forecasting, AttenGluco is evaluated under three protocols: isolated subject training, cohort-wise fine-tuning, and continual-learning or forgetting analysis. In isolated subject training, with $x_{wi}$0 train and $x_{wi}$1 test per subject and model reinitialization for each subject, AttenGluco improves RMSE, MAE, and correlation in every cohort relative to the multimodal 1D-CNN + LSTM baseline. The reported RMSE changes are healthy $x_{wi}$2, prediabetes $x_{wi}$3, oral T2DM $x_{wi}$4, and insulin T2DM $x_{wi}$5. The corresponding MAE changes are $x_{wi}$6, $x_{wi}$7, $x_{wi}$8, and $x_{wi}$9, while correlation rises from $x \in \mathbb{R}^d$0, $x \in \mathbb{R}^d$1, $x \in \mathbb{R}^d$2, and $x \in \mathbb{R}^d$3. In cohort-wise fine-tuning, the reported RMSE improvements are $x \in \mathbb{R}^d$4 for healthy, $x \in \mathbb{R}^d$5 for prediabetes, $x \in \mathbb{R}^d$6 for oral T2DM, and $x \in \mathbb{R}^d$7 for insulin T2DM. Across horizons, RMSE increases for both models, but AttenGluco’s increase is slower, especially at 30 and 60 minutes; in the insulin cohort, for example, the 30-minute RMSE changes from $x \in \mathbb{R}^d$8 and the 60-minute RMSE from $x \in \mathbb{R}^d$9. The abstract summarizes overall improvement as about $d = 12$0 in RMSE and $d = 12$1 in MAE relative to the multimodal LSTM baseline (Farahmand et al., 14 Feb 2025).

The continual-learning experiment in the same study probes catastrophic forgetting by sequentially introducing healthy, prediabetes, oral, and insulin cohorts without reinitialization. The reported result is qualitative rather than tabular: adding new cohorts degrades performance on prior cohorts for both models, indicating forgetting, but AttenGluco exhibits lower error rates than the baseline after fine-tuning on new cohorts. Specific per-transition RMSE, MAE, or correlation deltas are not reported (Farahmand et al., 14 Feb 2025).

For stress prediction and counterfactual generation, SenseCF evaluates GPT-4o in zero-shot and three-shot modes against DiCE, NICE, and CFNOW. On AI-READI, zero-shot GPT-4o yields validity $d = 12$2, distance $d = 12$3, sparsity $d = 12$4, and plausibility $d = 12$5, while few-shot GPT-4o yields validity $d = 12$6, distance $d = 12$7, sparsity $d = 12$8, and plausibility $d = 12$9. The baseline methods report DiCE: validity $L_2$0, distance $L_2$1, sparsity $L_2$2, plausibility $L_2$3; NICE: validity $L_2$4, distance $L_2$5, sparsity $L_2$6, plausibility $L_2$7; and CFNOW: validity $L_2$8, distance $L_2$9, sparsity $[0,1]$0, plausibility $[0,1]$1. The paper interprets the larger distances of LLM-generated counterfactuals as a limitation, but also reports that these samples improve downstream classifier accuracy when used for augmentation. Random Forest accuracy rises from $[0,1]$2 without augmentation to $[0,1]$3 using both zero-shot and few-shot CFs; XGBoost improves from $[0,1]$4 to $[0,1]$5; SVC from $[0,1]$6 to $[0,1]$7; and the neural network from $[0,1]$8 to $[0,1]$9 in the best reported setting. The abstract summarizes an average accuracy gain of $m = 1$0 (Soumma et al., 7 Jul 2025).

For synthetic data generation, RLSyn compares a PPO-based generator against EMR-WGAN and EHRDiff on the AI-READI tabular view. In privacy evaluation via membership inference AUROC, real data produce $m = 1$1 as a sanity check, EMR-WGAN yields $m = 1$2, RLSyn $m = 1$3, and EHRDiff $m = 1$4. In synthetic-to-real utility for T2D classification, the reported AUCs are $m = 1$5 for RLSyn, $m = 1$6 for EHRDiff, and $m = 1$7 for EMR-WGAN. In fidelity, real-to-synthetic AUC is $m = 1$8 for RLSyn, $m = 1$9 for EHRDiff, and the real-to-real baseline is $6$0. The authors interpret the relation $6$1 for EHRDiff as evidence of overfitting or replication tendencies on the smaller AI-READI dataset. For latent-space similarity and marginal fidelity, RLSyn is best on NMI $6$2 and DWD $6$3, whereas EHRDiff is best on CWC $6$4; EMR-WGAN is substantially worse on these fidelity measures. The paper’s stated conclusion is that, on AI-READI, RLSyn best balances privacy, utility, and fidelity in the small-sample regime (Espinosa-Dice et al., 24 Dec 2025).

6. Methodological limits, access, and future research uses

Several limitations recur across current uses of AI-READI. Missingness is explicit in wearable activity streams, where gaps arise from device recharging, and the forecasting study highlights mismatched temporal resolutions between CGM and activity data. At the same time, important preprocessing details are often underspecified: the interpolation method and normalization scheme in AttenGluco are not detailed, SenseCF does not specify stress-label thresholds or class distribution, and RLSyn does not report compute resources, training time, or random seeds. These omissions are not unique to AI-READI, but they materially affect reproducibility and cross-paper comparability (Farahmand et al., 14 Feb 2025, Soumma et al., 7 Jul 2025, Espinosa-Dice et al., 24 Dec 2025).

Access is described as public. One paper references the dataset as AI-READI: “Flagship Dataset of Type 2 Diabetes from the AI-READI Project (2.0.0) [Data set], FAIRhub,” with DOI https://doi.org/10.60775/fairhub.2, and cites an AI-READI program overview in Nature Metabolism with DOI https://doi.org/10.1038/s42255-024-01165-x. However, the arXiv studies do not describe licensing terms, registration procedures, data use agreements, or specific usage restrictions, and they direct readers to the FAIRhub entry for current access requirements (Farahmand et al., 14 Feb 2025).

The literature also identifies several forward directions. One line is broader modality integration: heart rate or stress index, environmental factors, sleep, and clinical variables are present in AI-READI but were omitted from AttenGluco’s forecasting model. Another is robust handling of missingness and irregularity through principled imputation, probabilistic modeling, or self-supervised learning, especially for cross-sensor gaps. A third is personalization under cohort shift, including meta-learning, transfer learning, or domain adaptation. Continual learning is particularly salient because sequential training across healthy, prediabetes, oral T2DM, and insulin cohorts reveals catastrophic forgetting; the paper explicitly points to replay-based methods, regularization-based approaches such as EWC, and modular architectures as possible remedies. In the counterfactual setting, best practices proposed in the literature include strict locking of immutable features, explicit enforcement of in-range plausibility, addition of physiologic coupling constraints, and human-in-the-loop clinical review for intervention-oriented counterfactuals (Farahmand et al., 14 Feb 2025, Soumma et al., 7 Jul 2025).

Taken together, these studies present AI-READI as a multimodal, multi-cohort testbed whose value lies in its breadth rather than in a single canonical benchmark instantiation. The dataset supports streaming and tabular formulations, near-term and longer-horizon prediction, counterfactual explanation, and privacy-sensitive synthesis. This suggests that its primary scientific role is to expose models to real-world challenges—missingness, irregular sampling, cohort heterogeneity, and representation shifts—while preserving sufficient modality diversity for future multimodal fusion and robustness research (Farahmand et al., 14 Feb 2025, Soumma et al., 7 Jul 2025, Espinosa-Dice et al., 24 Dec 2025).