H&E to recurrence score: A step forward, but not yet a substitute for genomic testing Authors: Shiqiong Zhou, Qinghua Ke Published: 2026 ABSTRACT Highlights•AI model predicts Oncotype DX recurrence score from H&E slides with AUC 0.898.•Recapitulates chemotherapy benefit patterns from the TAILORx trial.•Intratumoural heterogeneity and borderline predictions raise clinical uncertainty.•Node-positive and menopausal status require further validation (e.g., RxPONDER).•Tool is complementary, not a replacement for genomic testing. Genomic assays such as Oncotype DX (ODX) have transformed management of hormone receptor‑positive, HER2‑negative early breast cancer, but limited access due to cost and logistics persists globally. The motivation for developing an AI-based surrogate stems from the pressing need to extend precision oncology to underserved populations where genomic testing is unavailable, yet the clinical consequences of over- or under-treatment remain substantial. Shamai and colleagues developed a multimodal deep‑learning model that predicts ODX recurrence scores from routine H&E slides and clinicopathological variables, validated across the randomised TAILORx trial and six external cohorts including over 5000 patients [1]. The model achieved an AUC of 0.898 for identifying recurrence score ≥26 in the TAILORx‑test set, and recapitulated the prognostic and predictive patterns of the genomic assay, including chemotherapy benefit in premenopausal high‑risk patients and lack thereof in postmenopausal low‑risk patients. Notably, 31% of clinically high‑risk postmenopausal women were downgraded by AI to low risk, suggesting potential to reduce overtreatment. This is a significant step toward democratising precision oncology, but several issues need resolution before clinical adoption. Beyond the clinical uncertainties discussed below, the economic and infrastructural dimensions merit careful consideration. The AI tool may still offer value as a triage mechanism even with its current limitations—for example, by identifying patients with unequivocally low or high AI-predicted scores, thereby reserving costly genomic testing for intermediate or discordant cases. In resource-limited settings, such a triage could reduce overall healthcare costs, provided that the digital pathology infrastructure (slide scanners, storage, computational resources) is more feasible than genomic testing platforms. However, this assumption is not universally valid; in many low- and middle-income countries (LMIC), whole-slide scanners and high-performance computing remain scarce, whereas centralised genomic reference laboratories may already exist for infectious disease diagnostics. Comparative cost-effectiveness analyses, including capital and maintenance costs, are urgently needed before advocating for AI as a cost-saving alternative. Furthermore, the model still requires accurate IHC for ER, PR, and HER2 to define the eligible population [2]. in settings where these tests are unavailable or poorly standardised, the model could be applied incorrectly. A formal head-to-head comparison with conventional variables such as tumour grade and Ki‑67—which are cheap and widely available—would help clarify whether the AI approach offers added discrimination in resource-poor settings. Integrating AI-based molecular subtype classifiers directly from H&E is promising but requires independent validation [3,4]. and implementation science studies coupled with real-world LMIC cost-effectiveness analyses are necessary before claiming broad accessibility. First, how should intratumoural heterogeneity be managed in clinical decision‑making? The authors leverage TAILORx to assess chemotherapy benefit, but the trial's randomisation only included patients with recurrence scores 11–25 [5]. For patients with AI‑predicted scores <16 or ≥26, chemotherapy benefit estimates rely on simulated addition of small discordant subsets (1.4–1.8% of patients). The sensitivity analysis assumes that chemotherapy effect in these discordant groups mirrors the main population. However, patients with extreme discordance (AI score far from genomic score) may represent biologically distinct subgroups—for example, those with marked intratumoural heterogeneity. The authors themselves report that within‑patient variance in AI scores is significantly higher in outliers (p = 0.0055), and discordant cases often show pronounced heterogeneity. In such patients, which tissue block should guide treatment? The model aggregates predictions across multiple slides, but clinical practice requires a single decision. If AI and genomic scores conflict, and heterogeneity is the driver, then neither may represent the true biology of the entire tumour. A practical step forward would be to test pre‑specified sampling protocols prospectively—for example, using the block with the highest tumour cellularity or the most representative grade. For a postmenopausal patient with clinically high‑risk features, an AI low‑risk prediction (as in the 31% downgraded) but with high within‑patient variance, would clinicians confidently withhold chemotherapy? Prospective validation with defined sampling protocols is needed. Second, does the model's predictive value for chemotherapy benefit extend to patients with ambiguous or treatment‑altered menopausal status? TAILORx used age ≤50 as a surrogate for premenopausal, but in routine practice, menstrual status, hormone levels, and use of LHRH agonists vary considerably. The model's predictive value for chemotherapy benefit was derived directly from TAILORx's categorisation. For a 48‑year‑old patient with chemotherapy‑induced amenorrhoea or on LHRH agonists, does the AI's "premenopausal" chemotherapy benefit estimate apply? Similarly, for a 52‑year‑old with ongoing menses, would the model default to postmenopausal? No current AI model can resolve this unless the original training data included hormonal or menstrual details, which it did not. To address this, the model could potentially be retrained or fine-tuned on datasets that capture granular menopausal status (e.g., serum oestradiol, follicle-stimulating hormone, and documented LHRH agonist use) to refine predictions. In the interim, combining the AI prediction with clinically documented menopausal status—when available—could improve decision-making: for instance, a clinician might override the model's default based on actual menstrual history. The practical frequency of ambiguous menopausal status in the target population is non-negligible; many breast cancer patients are diagnosed in the perimenopausal age range (45–55 years), and chemotherapy-induced amenorrhoea is common. Future validation datasets should collect granular menstrual history and LHRH agonist use to enable more precise subgroup analyses. Third, can the model safely guide treatment in node‑positive disease? The authors emphasise generalisability to node‑positive patients, yet all training and randomised validation came from node-negative TAILORx. The RxPONDER trial showed that postmenopausal node‑positive patients with recurrence score <26 do not benefit from chemotherapy, whereas premenopausal node‑positive patients do, irrespective of score [6]. The current model's performance in node‑positive external cohorts (AUC ∼0.87) is encouraging, but the critical predictive question—whether a postmenopausal node‑positive patient with AI low risk can safely avoid chemotherapy—remains untested in a randomised dataset. The authors note that only 3.4% of AI low‑risk node‑positive patients had recurrence score ≥26, and 0.6% had ≥31, suggesting low risk of undertreatment. However, in RxPONDER, the hazard ratio for chemotherapy in postmenopausal node‑positive patients with recurrence score <26 was 1.02 (95% CI 0.82–1.26), consistent with no benefit. Even a small fraction of patients with true recurrence score ≥26 misclassified as AI low risk could alter the risk‑benefit calculus if those patients do derive benefit. Retrospective validation using archived RxPONDER slides would be sufficient to answer the predictive question and should be pursued before adopting the model for node‑positive decision‑making. Moreover, the integration of this AI tool into existing clinical pathways requires practical considerations: expected turnaround time (likely within hours to days, versus 7–14 days for genomic testing) could facilitate faster decision-making, but the interpretation of AI predictions would need a clear governance structure—pathologists are best positioned to assess slide quality and sampling, while oncologists should integrate the AI output with clinical factors. Regulatory approvals (e.g., FDA clearance or CE marking) and certification of digital pathology laboratories are prerequisites for routine use, and these processes are not trivial, especially in LMIC. Beyond these clinical uncertainties, calibration and threshold uncertainty remain practical barriers. The calibration method, though clever (requiring only the mean ODX score of the target cohort via clinical variables), assumes that the relationship between AI predictions and true recurrence scores is linear and consistent across the risk spectrum. The authors show improved root mean squared error after calibration, but clinical consequence of miscalibration in the tails (e.g., predicted score 24 vs. 26) could be substantial. For a premenopausal patient, a predicted score of 25 (AI intermediate) versus 26 (AI high risk) changes the recommendation from "no evidence for chemotherapy benefit" to "chemotherapy recommended." Confidence intervals around individual predictions are not provided. AI predictions near thresholds will inevitably have classification uncertainty; without a measure of confidence (e.g., prediction interval or probability of crossing the threshold), clinicians may struggle with borderline cases. To quantify this, the proportion of patients with predictions within a narrow zone around each clinically relevant cutoff (e.g., 22–28 for the ≥26 threshold, and 14–18 for the <16 threshold) can be estimated from the TAILORx-test set; based on reported distributions, approximately 15–20% of patients fall into these grey zones. For such patients, genomic testing should be strongly recommended regardless of the AI prediction. We propose defining a "zone of uncertainty" (e.g., predicted score 22–28) where the AI prediction is deemed non-decisive, and where the gold-standard assay remains mandatory. This zone could be refined through prospective calibration studies that provide prediction intervals for individual outputs. Until such uncertainty metrics are available, clinicians should exercise caution in borderline cases. Finally, the future work outlined by the authors—while promising—remains underdeveloped. We propose three concrete directions: (1) prospective, multicentre implementation trials that embed pre‑specified sampling protocols to address heterogeneity, with predefined rules for resolving AI-genomic discordance; (2) retrospective validation using the archived RxPONDER trial slides to confirm predictive value in node‑positive disease, with separate analyses by menopausal status and LHRH agonist use; and (3) formal cost‑effectiveness and infrastructure feasibility studies in LMIC settings, directly comparing the AI approach with standard clinical variables (grade, Ki‑67) and with genomic testing, alongside implementation science to evaluate workflow integration, turnaround times, and user acceptability. Additionally, the model's open-source code and foundation-model architecture provide a strong basis for external fine‑tuning; collaboration with global consortia (e.g., the International Breast Cancer Study Group) could accelerate these validations. Despite these limitations, Shamai and colleagues have delivered a landmark study: the first AI pathology model for recurrence score prediction validated using randomised trial data to assess chemotherapy benefit. The extensive external validation, use of a foundation model for generalisability [7]. and open‑source code release set a new standard. For practising oncologists today, the message is not to replace genomic testing but to view this AI tool as a complementary decision aid—particularly where ODX is unavailable. For patients with clear-cut AI low‑risk (e.g., postmenopausal, AI <16, and low clinical risk) or high‑risk (premenopausal, AI ≥26, and high clinical risk), the data support treatment decisions with reasonable confidence. For the substantial intermediate group (42% in TAILORx‑test) and for discordant or borderline cases, genomic testing remains the gold standard. Three critical questions must be addressed next: first, how should intratumoural heterogeneity be managed through pre‑specified sampling protocols? second, can the model be validated in randomised node‑positive cohorts (e.g., retrospective RxPONDER) to confirm predictive value? third, what is the performance and cost‑effectiveness when deployed prospectively in low‑resource health systems, and how do standard variables like grade and Ki‑67 compare? Answering these will transform this impressive proof‑of‑concept into a globally implementable tool that truly democratises precision oncology. Ethical statement As this article is a commentary on published clinical trial data and does not involve original research with human participants, patient data, or animal subjects, no ethical approval was required. Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. CRediT authorship contribution statement Shiqiong Zhou: Writing – review & editing, Writing – original draft, Conceptualization. Qinghua Ke: Writing – review & editing, Writing – original draft, Conceptualization. Declaration of competing interest We declare no competing interests. All authors had full access to the complete data of the study and took final responsibility for the decision to submit the manuscript for publication. Data and codes are available from the corresponding author upon reasonable request.