GraphBEAT: a temporal patient-record graph to estimate pancreatic, liver and intestinal cancer risk 12 to 36 months before diagnosis, validated within a single health system
A Mayo Clinic team proposes GraphBEAT, a model that turns a patient's history of diagnoses and blood tests into a temporal graph to estimate the risk of pancreatic, liver or intestinal cancer 12, 24 or 36 months before diagnosis. On 201,322 patients from four Mayo sites, it reaches an AUROC of 0.919 for pancreatic cancer at 12 months, ahead of other graph networks by 0.004 to 0.009 AUROC and ahead of a Transformer and an LSTM by 0.07 to 0.10, with roughly 20 times fewer parameters than the latter two. The result is interesting as a first-stage screening step, but this is a preprint, from a single health system, with cases defined by billing codes and probabilities whose calibration has not been studied.
Background
Pancreatic cancer is often diagnosed late. Yet electronic health records hold early traces: recent-onset diabetes, weight loss, a liver abnormality, drifting blood tests. Using them is hard because these records are irregular (not everyone is measured on the same things), sparse and incomplete. Common approaches flatten the history into a sequence handled by an LSTM (a recurrent network that reads events in order) or a Transformer (an attention-based architecture that weighs each event against the others), both computationally heavy.
GraphBEAT, posted to arXiv on 15 October 2026 by Leo Li-Han, Hojjat Salehinejad and colleagues from the Department of Surgery and the Kern Center at Mayo Clinic (Rochester), proposes another representation: a graph, that is, a set of nodes (here clinical events) linked by edges. The text has not been peer reviewed. We found neither a funding statement nor a conflict-of-interest statement in it.
The method
Data. De-identified electronic health record data from the Mayo Clinic Platform, from 1985 to 2024, with up to 20 years of history per patient: ICD-9/10 diagnosis codes truncated to three characters (2,698 codes) and 31 prespecified blood tests, a vocabulary of 2,729 items. The cohort has 201,322 patients across four sites: Rochester (93,558), the 17-hospital community network Mayo Clinic Health System, MCHS (57,457), Arizona (25,418) and Florida (24,889). Cases are 6,017 pancreatic cancers (stated prevalence 3.0%), 5,330 liver cancers (2.6%) and 13,043 intestinal cancers (6.5%), all compared with one group of 177,081 controls matched on age and sex. At least three years of history and three laboratory measurements are required.
The graph. The history is cut into 30-day buckets. Each patient gets a graph in which a node represents an item observed in a bucket (a code, or a lab test with its mean value), linked to every node of the next bucket. Each node sums three pieces of information: the item's identity, its value and its position in time. Messages flow from one bucket to the next with an exponential decay depending on the time gap (message passing: each node updates its representation from its predecessors), then attention pooling condenses everything into one patient vector. The model has 0.361 million parameters.
Validation. Leave-one-site-out: MCHS, Arizona and Florida are each held out in turn for testing, while Rochester always stays in training. Ten models per setting (ten-fold cross-validation), with mean and 95% confidence interval. Controls are downsampled to 1 for 3 during training. Metrics are the AUROC (the probability that a patient with cancer gets a higher score than one without) and AP, average precision, which summarises the trade-off between sensitivity and precision and depends heavily on prevalence. Comparators are three graph networks (GraphSAGE, GAT, GCN), a Transformer and an LSTM.
The results
ML metrics. Pancreatic cancer at 12 months: AUROC 0.919 [0.914–0.924] and AP 0.523 for GraphBEAT, versus 0.913 and 0.499 for GraphSAGE, 0.837 and 0.341 for the Transformer, 0.808 and 0.264 for the LSTM. At 36 months, GraphBEAT keeps an AUROC of 0.862 and an AP of 0.331. For liver, the AUROC goes from 0.938 at 12 months to 0.895 at 36 months; for intestine, from 0.900 to 0.852. Against the best graph network, the gain is 0.004 to 0.009 in AUROC and 0.024 to 0.042 in AP; against the best sequence model, 0.070 to 0.102 in AUROC and 0.125 to 0.202 in AP.
Compute cost. On the 12-month pancreatic task, GraphBEAT needs 12.5 million operations (FLOPs) per patient versus 3,995 million for the Transformer and 3,231 million for the LSTM, and 0.361 million parameters versus 8.3 and 7.2 million.
Ablations. On the MCHS site, pancreatic cancer at 12 months, the AUROC is 0.936 with identity, value and time; 0.922 without time; 0.712 with item identity alone and 0.549 with time alone. Laboratory values carry most of the signal.
Clinical translation. On MCHS, the authors report a net benefit of 0.018, 0.016 and 0.015 at 12, 24 and 36 months for pancreatic cancer, at a threshold of 0.03, or about 18, 16 and 15 net true-positive detections per 1,000 patients. Net benefit analysis weighs true positives against false positives, weighted by the chosen threshold. As an illustration (editorial calculation), a prevalence of 3.0% means about 30 cases per 1,000 patients in this cohort, and an AP of 0.523 is about 17 times the value of random ranking (0.03). But this prevalence is that of the constructed cohort, not of a screening population: if the disease were ten times rarer, the positive predictive value would be, at equal discrimination, roughly ten times lower, and the number of false positives per detected case much higher.
What is good
A large cohort and site-wise validation. 201,322 patients, four sites, up to 20 years of history, and testing on sites never seen in training, including 17 community hospitals, a different population from tertiary centres. The lead-time window (12, 24, 36 months) is excluded from the test data, which limits leakage of information about the imminent diagnosis.
Ablations that isolate what matters. Time alone predicts almost nothing (AUROC 0.549) and identity alone stays modest (0.712): the signal comes from the combination of what is observed, its value and its timing. This makes a simple visit-frequency effect less plausible, without ruling it out.
A light model and results shown with intervals. 0.36 million parameters and a per-patient cost several hundred times lower than sequence models: a concrete argument for deployment across millions of records. The main limitations (single system, cases defined by codes, calibration) are acknowledged by the authors.
What is less good
An incomplete comparator (biased comparator). The gain over GraphSAGE is 0.004 to 0.009 AUROC, of the same order as the width of the table's confidence intervals, which overlap (0.914–0.924 versus 0.907–0.919 for pancreatic cancer at 12 months), whereas the Transformer and LSTM are clearly outpaced, notably with an AUROC of 0.760 for the Transformer on intestine, behind the LSTM (0.830): this suggests lightly tuned baselines. In the results we read, there is no logistic regression or boosted trees on aggregated variables, and no published pancreatic risk score.
An "external" validation that is not (population bias, possible leakage). The three test sites belong to the same system, with the same electronic record, and Rochester is always in training. Cases are defined by diagnosis codes alone, without a registry. Controls have an index date set one year before their last encounter, whereas for cases it is the first cancer code: the two groups are thus not observed the same way, and the requirement of three years of history and three lab tests selects heavily followed patients. These choices can create shortcuts (shortcut learning) that the ablations do not entirely exclude.
Flattering metrics (misleading metric) and no calibration. AP depends on prevalence, which is artificial here (3.0%, matching, controls downsampled to 1 for 3). The authors acknowledge that downsampling changed the class priors and that calibration must be studied before any clinical use: the 0.03 threshold of the net benefit analysis therefore rests on probabilities whose accuracy is unknown. The ablations are reported on the MCHS test set itself. Finally, code is not mentioned and the data cannot be shared.
What it changes
For the research community. An argument for temporal graph representations for irregular records, at a much lower compute cost. The logical next step is a comparison against strong tabular models and published scores, in an outside health system, with calibration and registry-confirmed cases.
For clinicians. Nothing in practice today: this is a retrospective preprint. In the medium term, the use the authors envisage is first-line screening ahead of targeted tests. The decisive question will be how many patients must be worked up to find one cancer, in a real population rather than a case-control cohort.
For patients and the public. No tool is available. A risk score drawn from a medical record does not say that one has cancer: it can help decide who to examine more closely, and its value depends on how rare the disease is.
Going further
Preprint: Li-Han L. Y., Varghese C., Habermann E. B., Truty M. J., Perry W. R. G., Thiels C. A., Salehinejad H., "GraphBEAT: Graph-based Bucketized Event Aggregation Over Time for Early Risk Stratification of Gastrointestinal Cancers", arXiv:2610.22014, 15 October 2026 (CC BY-NC-ND 4.0 licence). Comparators: GraphSAGE, GAT, GCN, Transformer, LSTM. Data: Mayo Clinic Platform, not shareable; Mayo Clinic Institutional Review Board approval.
Editorial transparency: this article was produced with the help of artificial intelligence tools from the preprint; figures come from the paper, except conversions and comparisons calculated by the editorial team. The English, Spanish and Chinese versions are machine translations.