Agentic discovery of blood biomarker from distilled private health records

1 Introduction

• The paper addresses the privacy and evaluation gap by releasing a scorer that lets external agents rank CBC-expression hypotheses using evidence distilled from private records.CBC arithmetic summaries are attractive for their low cost, interpretability, and broad availability, while manually enumerating the 26 trillion possible combinations is intractable.Source: §1 · pp. 1–3

2.1 A private cohort was distilled into a released scorer for CBC expression trees

• The released disease-specific GATs predicted held-out expression-AUC rankings with median Spearman correlation ≈0.93 across 13 diseases.Each model was trained inside the Clalit boundary on expression–AUC pairs, not patient-level records; the corpus used about 2 million synthetic expressions per disease. Maximum predicted-AUC MAE was 0.0186, and NDCG@5000 ranged from 0.984 to 0.999.Source: §2.1 · p. 3, Figure 2 · p. 4, Table 2 · p. 5

2.2 Agent-discovered expressions did not uniformly generalize across three external validation cohorts

• External transfer varied by cohort: discovered expressions matched or exceeded starting-expression AUC in 12 of 12 evaluable MIMIC-IV comparisons, but results were weaker on EHRShot and mixed on NHANES.On EHRShot, 7 of 12 evaluable comparisons improved, but only 1 met the strict confidence-interval criterion. In NHANES, rheumatoid arthritis AUC rose from 0.522 to 0.640 with separated confidence intervals, while psoriasis AUC fell from 0.592 to 0.570 with overlapping intervals; these analyses used the two verified-diagnosis phenotypes.Source: §2.2 · pp. 4–5, Table 3 · p. 5, Figure 3 · p. 9

2.3 The released GAT improved the ranking of leading LLM-generated biomarker candidates

• The scorer-ranked expression beat the first expression from external LLM tools in 67.9% of 81 verified-diagnosis comparisons, with a median gain of +1.50 AUC pp.The advantage was cohort-dependent: it was observed on MIMIC-IV, suggested but not established in the small verified-diagnosis NHANES panel, and absent on EHRShot. Across tools, scorer wins were 74.1% for GPT Deep Research, 66.7% for Gemini 3 Pro, and 63.0% for SciSpace BM Agent.Source: §2.3 · p. 6

2.4 A single-CBC-feature baseline on the development cohort

• On the held-out development split, scorer-ranked candidates exceeded the best single CBC feature in every disease, by a mean of +2.83 AUC pp.The best single feature exceeded the best first-emitted tool candidate in 8 of 13 diseases and led by +1.80 AUC pp averaged across all 13; values were selected on the reported split and are optimistic relative to a marker fixed in advance.Source: §2.4 · pp. 6–7, Table 5 · p. 8

2.5 Prediction-guided search improved post-hoc AUC

• Across 13 diseases, prediction-guided search improved post-hoc realized AUC by a median of +4.18 AUC pp over the starting expression.The gain ranged from +1.89 AUC pp for celiac disease to +15.92 AUC pp for SLE; validation AUC was measured only after selection and did not guide ranking or stopping. Familial Mediterranean fever’s estimate was unstable; excluding it, the median gain was +4.08 AUC pp and the mean was +5.95 AUC pp. A median of five iterations captured 95% of predicted-score gain.Source: §2.5 · pp. 7–8, Table 6 · p. 8, Figure 4 · p. 10

2.6 A random-scoring control isolates the scorer’s contribution

• Replacing the trained scorer with random scores reduced mean post-hoc gain from +5.84 to +0.07 AUC pp, with only 4 of 13 control runs nominally positive versus 13 of 13 GAT-guided runs.Excluding familial Mediterranean fever, the mean gain was +5.95 versus +0.13 AUC pp; control starting expressions were independently re-derived and differed from the main arm in 8 of 13 diseases, so the arms were not paired at the starting point.Source: §2.6 · pp. 9–10, Table 8 · p. 10

2.7 Shapley attribution shows the discovered expressions largely recover known blood count biology

• End-to-end Shapley attribution assessed how each CBC feature functioned within an expression, classifying signed contributions as Expected, Surprising, or Unclear against clinical literature.The literature review used multisource review and adversarial citation verification, discarding unverifiable references.Source: §2.7 · p. 11, Figure 5 · p. 12
• Surprising contributions clustered in several lower-AUC expressions, but concordance did not track AUC monotonically.Crohn’s disease and celiac disease were fully expected despite low AUC; higher mean corpuscular volume in Sjögren’s syndrome and systemic sclerosis may reflect treatment or age, while feature reuse across fraction bars suggested symbolic-regression overfit.Source: §2.7 · p. 11, Table 9 · p. 11
• The analysis provides a literature-grounded audit layer that identifies high-AUC expressions consistent with inflammatory-anemia physiology and components of weaker expressions requiring scrutiny before downstream use.Per-disease explainers with feature-level verdicts and verified citations are provided as Supplementary Data 1.Source: §2.7 · p. 11
• 41 of 64 contributions were Expected (21 Surprising, 2 Unclear); lupus (AUC 0.676) was fully expected, and rheumatoid arthritis (AUC 0.615) had five of six expected components.Expected contributions chiefly reflected the anemia-of-inflammation pattern: higher red-cell-distribution-width and lower hemoglobin and red-cell terms raised predicted risk, alongside disease-specific leukocyte directions.Source: Table 9 · p. 11, §2.7 · p. 11

3 Discussion

• The GAT weights provide a portable scorer trained on expression-level AUC labels inside the CHS boundary, without releasing patient-level records.External users submit a disease identifier and CBC expression to receive a predicted AUC score.Source: §3 · pp. 13–14
• Across 81 verified-diagnosis comparisons, reranking won 67.9% [57.1%,77.1%] with median paired gain +1.50 AUC pp; rates were 86.1% in MIMIC-IV and 48.7% in EHRShot.The EHRShot comparison was not statistically significant (p = 0.92), so the advantage was cohort-dependent.Source: §3 · p. 13
• The median realized-AUC improvement was +4.18 AUC pp across 13 disease runs, a selection-bound validation yield rather than autonomous clinical gain.Excluding the unstable familial Mediterranean fever cell, the median improvement was +4.08 AUC pp across 12 runs.Source: §3 · p. 13
• Replacing the GAT with a uniform random tool reduced mean post-hoc gain across 13 diseases from +5.84 to +0.07 AUC pp.The loop, prompts, feature allowlist, and external-validation pipeline were held fixed.Source: §3 · p. 13
• Improvement depended on the starting expression: near-chance baselines left more room for arithmetic gains, while stronger ratios left less headroom.Most starting expressions were NLR or a related CBC ratio, already known to carry inflammatory signal.Source: §3 · p. 13
• Only one of 13 runs completed the 15-iteration budget; ten would require further iterations before their outcomes could be interpreted.The audit classes describe search depth and comparison against independent tools, not improvement over each run’s starting expression.Source: §3 · p. 13
• On the primary cohort, the discovered expression scored below the strongest expression in the independent-tool comparator set: a 15-iteration blinded search did not beat the best of roughly 200 expressions combined from three tools.Source: §3 · p. 14
• Future extensions include multi-institution ensembles, longitudinal CBCs, other routine panels, more diseases, and a single GAT for multiple diseases.The study’s current disease panel covers 13 conditions.Source: §3 · p. 14

4.1 Study design and terminology

• For 13 diseases, the team trained GAT scorers on synthetic CBC-expression trees and private Clalit AUC labels, then released only model weights and non-patient metadata.Each disease’s synthetic corpus contained about two million trees built from 14 CBC features and four arithmetic operators, capped at depth three; the six-layer GAT returned predicted AUC scores for candidates without querying patient records.Source: §4.1 · p. 14, §4.3 · p. 15, §4.4 · p. 16

4.2 Cohorts, phenotypes, and CBC measurements

• Cases and controls were balanced 1:1 using pre-diagnosis CBCs, and fixed expressions were evaluated without retraining in MIMIC-IV, EHRShot, and verified-diagnosis NHANES phenotypes.The case CBC was the last measurement within five years before diagnosis, with a 30-day blanking window; primary NHANES validation covered psoriasis and rheumatoid arthritis only. MIMIC-IV, EHRShot, and NHANES contributed different disease panels.Source: §4.2 · pp. 14–15, §4.9 · p. 18

4.3 CBC expression space

• Candidates were binary trees over 14 CBC features and four arithmetic operators, while the read-only scorer API returned predicted-AUC scores without accepting patient fields or returning patient-derived records.Numeric literals were prohibited and candidates were canonicalized to deduplicate equivalent commutative forms; the intended maximum depth was three.Source: §4.3 · p. 15, §4.5 · pp. 16–17

4.6 LLM-guided propose–score–refine search

• Five selected expressions exceeded the depth-three training grammar, so their predicted scores were extrapolations; agents saw scorer scores, not external AUCs, during search.The validity filter did not enforce the depth bound, and the scorer returned ordinary range-bounded scores for over-depth trees without an out-of-distribution warning. The search began from public-literature seeds, proposed up to 200 candidates per iteration, and ran 5–15 iterations; external evaluation occurred after candidate selection.Source: §4.6 · p. 17

4.7 Post-hoc audit of iterative search runs

• In all 10 of 13 runs with located comparator sets, the discovered expression scored below the strongest independent-tool comparator; three recorded gate passes came only from the default rule.The post-hoc audit used a strict point-estimate comparison without confidence intervals, and its classes described search depth and comparator status rather than improvement over each run’s starting expression.Source: §4.7 · p. 18

4.8 Random-scoring ablation

• The random-scoring control held the search setup fixed but used different starting ratios in 8 of 13 diseases, and random scores were effectively uncorrelated with realized AUC.Each arm’s gain was measured against its own starting expression; most control runs reached the 15-iteration cap.Source: §4.8 · p. 18

4.10 External LLM-tool reranking benchmark

• The reranking benchmark compared each tool’s first-emitted candidate with the scorer-selected candidate on realized external AUC, using 81 primary cells and a separate 9-cell NHANES sensitivity panel.For the single-feature baseline, columns used maxima over 14 CBC features, each tool’s first five candidates, or a scorer-ranked pool of 25 candidates; a null simulation estimated the selection effect of the larger pool, not an exact correction.Source: §4.10 · pp. 18–19, §4.11 · p. 19

4.12 Shapley attribution and literature-concordance protocol

• Exact interventional Shapley values attributed each expression’s MIMIC-IV prediction to its component CBC features, but the audit was interpretive rather than causal and could reflect acuity, treatment, or age confounding.The logistic model used the whole expression as its predictor, with the cohort median as reference; ambulatory re-attribution remained open because EHRShot had a median of only 47 positive cases per disease.Source: §4.12 · pp. 19–20

4.13 Privacy analysis

• The scorer training corpus contains about two million candidate expressions per disease labeled with direction-agnostic case–control AUCs, not patient-level data, and the exported artifact contains only trained weights and non-patient metadata.Although trained models can in principle leak information, even recovering every training pair would reveal only expressions and cohort-level AUCs, not patients’ laboratory values, diagnoses, dates, or identities.Source: §4.13 · p. 20
• A patient’s replacement shifts any label by at most about one part in a thousand, versus per-disease scorer mean absolute error of 0.0030–0.0186.Each label is an AUC over a balanced case–control subcohort of 2,390–111,835 patients (Table 1), making the shift many times smaller than the scorer’s approximation error (Table 2).Source: §4.13 · p. 20
• The scorers reveal which CBC expressions separate cases from controls, but not cohort composition beyond public information, and an empirical membership-inference or extraction audit would probe an attack surface this design does not create.The case–control ratio is fixed at 1:1, disease prevalence never enters training, and subcohort sizes are disclosed in Table 1.Source: §4.13 · p. 20

Data availability

• MIMIC-IV is available to PhysioNet-credentialed researchers.Access is through PhysioNet.Source: Data availability · p. 20
• EHRShot access is through the Stanford Medicine Data Use Agreement, and cohort construction follows the EHRShot benchmark protocol.Source: Data availability · p. 20
• NHANES is public domain, but CHS individual-level data cannot be shared publicly because of national and organizational privacy regulations.Source: Data availability · p. 20

Ethics declarations

• The Clalit Community Clinics Helsinki committee approved the CHS component under approval number 1COM0161-19.Source: Ethics declarations · p. 21
• The study retrospectively analyzed de-identified electronic health records entirely within the CHS data boundary, without contacting participants.Source: Ethics declarations · p. 21
• No prospective or interventional procedure was performed, and CHS analyses followed the Declaration of Helsinki and applicable Israeli and CHS data-protection regulations.Source: Ethics declarations · p. 21

Supplementary Data 1 Discovered CBC biomarkers explained: Shapley attribution and literature concordance

• Each expression has a self-contained explainer with a glass-box tree whose blood-test leaves show Shapley contributions to disease prediction.Shapley values are computed out-of-sample on MIMIC-IV.Source: Supplementary Data 1 Discovered CBC… · p. 24
• Each explainer classifies every component as Expected, Surprising, or Unclear against the clinical literature and provides supporting evidence.Source: Supplementary Data 1 Discovered CBC… · p. 24
• The reports and references are provided as is and may contain errors; citations were resolved through PubMed and independently re-verified.The bibliography is shared across all twelve explainers.Source: Supplementary Data 1 Discovered CBC… · p. 24

Systemic lupus (SLE)

• Blood-test leaves encode Shapley impact by color (red raises risk; blue lowers it), with triangles showing direction and circles or stars marking expected or surprising components.Source: Systemic lupus (SLE) · p. 25
• The expression ((RBC + WBC) · HB) / (RDW + NEUT% + MONO%) has out-of-sample AUC (MIMIC) ≈0.6757.Source: Systemic lupus (SLE) · p. 25
• Higher WBC, HB, and RBC each lower predicted SLE risk, consistent with their expected reduction in the disease.The passages identify leukopenia as a core criterion-level feature and anemia as a common laboratory abnormality.Source: Systemic lupus (SLE) · p. 25
• Higher NEUT%, MONO%, and RDW each raise predicted SLE risk, consistent with their expected elevation in the disease.Source: Systemic lupus (SLE) · p. 25

• The expression (RDW - RBC) · (MCV + RDW - MCH) / (HB · MCHC - RDW) has an out-of-sample MIMIC AUC of approximately 0.6146.Source: Rheumatoid arthritis · p. 26
• Higher RDW raises predicted risk, consistent with elevated RDW and inflammation-related disruption of erythropoiesis and iron metabolism in rheumatoid arthritis.Source: Rheumatoid arthritis · p. 26
• Higher HB and RBC lower predicted risk, consistent with anemia of chronic disease and reduced red-cell counts in rheumatoid arthritis.Source: Rheumatoid arthritis · p. 26
• Higher MCV raises predicted risk, consistent with MCV being variably altered in rheumatoid arthritis.The passage notes two competing mechanisms.Source: Rheumatoid arthritis · p. 26
• Higher MCHC lowers predicted risk, a surprising direction because MCHC is typically variably altered and is not RA-specific.Source: Rheumatoid arthritis · p. 26
• Higher MCH lowers predicted risk, consistent with reduced MCH in active rheumatoid arthritis.Source: Rheumatoid arthritis · p. 26

• The expression (HB · HB · RBC) / (LYM% · PLT + HCT) has an out-of-sample MIMIC AUC of approximately 0.6056.Source: Hashimoto thyroiditis · p. 27
• Higher LYM% raises predicted risk, unexpectedly given that LYM% is typically reduced; the passage describes lymphocytes homing into the thyroid.Source: Hashimoto thyroiditis · p. 27
• Higher PLT raises predicted risk, consistent with elevated platelets in this chronic systemic-inflammatory autoimmune state.Source: Hashimoto thyroiditis · p. 27
• Higher HB and RBC lower predicted risk, consistent with lower hemoglobin and thyroid-hormone-related erythropoiesis in Hashimoto thyroiditis.Source: Hashimoto thyroiditis · p. 27
• Higher HCT raises predicted risk, surprisingly because HCT is typically reduced in Hashimoto thyroiditis, a leading cause of hypothyroidism.Source: Hashimoto thyroiditis · p. 27

• The expression (MCH · HB + HCT - RDW) · (HB + MCV · MCHC) has an out-of-sample MIMIC AUC of approximately 0.6015.Source: Hyperthyroidism · p. 28
• Higher HB lowers predicted risk, while the passage says the literature on hemoglobin in hyperthyroidism is limited.Thyrotoxicosis is described as tending to lower hemoglobin and raise anemia prevalence versus euthyroid controls.Source: Hyperthyroidism · p. 28
• Higher MCH, MCV, and MCHC lower predicted risk, consistent with reduced indices associated with altered erythropoiesis and iron handling.Source: Hyperthyroidism · p. 28
• Higher HCT lowers predicted risk, although thyroid-driven erythropoiesis can raise HCT while anemia-promoting effects dominate at the population level.The passage attributes the anemia-promoting arm to factors including hemodilution, altered iron metabolism, oxidative stress, red-cell turnover, and nutrient deficiencies.Source: Hyperthyroidism · p. 28
• Higher RDW raises predicted risk, surprisingly because RDW is typically elevated in hyperthyroidism; the passage associates RDW increases mainly with hypothyroidism and autoimmune thyroiditis.Source: Hyperthyroidism · p. 28

Psoriasis

• Psoriasis: (LYM% + RDW + RDW) / (HB · HCT · MCV · NEUT% · (BASO% + EOS%)) has an out-of-sample MIMIC AUC of approximately 0.5966.Higher MCV raises predicted risk, consistent with its modest elevation relative to healthy controls.Source: Psoriasis · p. 29
• Psoriasis: Higher EOS% and NEUT% raise predicted risk, while higher LYM% lowers it, consistent with the stated inflammatory-disease patterns.Source: Psoriasis · p. 29
• Psoriasis: Higher BASO% raises predicted risk despite typically reduced BASO%, a surprising direction for a disease described as Th1/Th17-driven rather than Th2/IgE/atopic.Source: Psoriasis · p. 29
• Psoriasis: Higher HB and HCT raise predicted risk, whereas higher RDW lowers it, despite the passages describing these measures as typically reduced or elevated, respectively.Source: Psoriasis · p. 29

Sjogren’s syndrome

• Sjogren’s syndrome: RDW · MCV - HB · (HCT + RBC) has an out-of-sample MIMIC AUC of approximately 0.5964; higher RDW and MCV raise predicted risk.Source: Sjogren’s syndrome · p. 30
• Sjogren’s syndrome: Higher HB, HCT, and RBC lower predicted risk, consistent with reduced measures and anemia of chronic disease described in the passages.The HCT passage also lists autoimmune hemolytic anemia, marrow involvement, and nutritional or renal contributions.Source: Sjogren’s syndrome · p. 30

Systemic sclerosis

• The systemic-sclerosis expression ((RDW · MCV) - HB · MCH) / (HB · MCHC) · RDW has out-of-sample AUC (MIMIC) ≈0.5875.Source: Systemic sclerosis · p. 31
• Higher RDW, MCH, and MCHC lower or raise predicted systemic-sclerosis risk as described, while higher MCV raises risk despite typically reduced MCV.Higher RDW raises risk; higher MCH and MCHC lower it. Higher HB also lowers risk, consistent with reduced HB in systemic sclerosis.Source: Systemic sclerosis · p. 31

Ulcerative colitis

• The ulcerative-colitis expression (PLT · RDW - LYM% · HB) · (RDW · (RDW + RDW + NEUT%)) / (LYM% · HB - MONO%) has out-of-sample AUC (MIMIC) ≈0.5715.Source: Ulcerative colitis · p. 32
• Higher LYM% and HB raise predicted ulcerative-colitis risk, whereas higher PLT, RDW, and NEUT% lower it, contrary to the stated typical patterns for these measures.MONO% raises predicted risk, consistent with its described elevation in ulcerative colitis.Source: Ulcerative colitis · p. 32

Crohn’s disease

• The Crohn’s-disease expression (PLT · RDW + NEUT% · RDW) / (LYM% · HB + LYM%) has out-of-sample AUC (MIMIC) ≈0.5697.Source: Crohn’s disease · p. 33
• Higher PLT, RDW, and NEUT% raise predicted Crohn’s-disease risk, while higher LYM% and HB lower it, consistent with the stated disease patterns.Source: Crohn’s disease · p. 33

Multiple sclerosis

• The multiple-sclerosis expression HB · RBC + LYM% - NEUT% has out-of-sample AUC (MIMIC) ≈0.5697.Source: Multiple sclerosis · p. 34
• Higher NEUT%, HB, and RBC raise predicted multiple-sclerosis risk despite typically reduced levels, while the LYM% direction is unclear against limited literature.Source: Multiple sclerosis · p. 34

Type 1 diabetes

• The type 1 diabetes expression (WBC + RDW + EOS%) / (LYM% + HB + HCT) has out-of-sample AUC (MIMIC) ≈0.5679.Source: Type 1 diabetes · p. 35
• Higher WBC, RDW, and EOS% raise predicted type 1 diabetes risk despite the stated typical patterns, while higher LYM%, HCT, and HB lower it.The passage describes higher WBC as surprising because elevated total WBC is not a recognized hallmark of type 1 diabetes.Source: Type 1 diabetes · p. 35

Celiac disease

• The celiac-disease expression (RBC - RDW) / (RBC + RDW) has out-of-sample AUC (MIMIC) ≈0.5569.Source: Celiac disease · p. 36
• Higher RBC lowers predicted celiac-disease risk and higher RDW raises it, consistent with the described anemia and nutrient-absorption patterns.Source: Celiac disease · p. 36