AI Histologic Grading of Breast Carcinoma: What "predGrade" Can and Cannot Do

The diagnostic and clinical question

Histologic grade is one of the oldest and most durable prognostic variables in breast pathology. The Nottingham histologic grade (NHG), also called the Nottingham Grading System or Elston–Ellis grade, combines three morphologic components — tubule/gland formation, nuclear pleomorphism, and mitotic count — into a composite score that separates patients into grade 1 (well differentiated), grade 2 (moderately differentiated), and grade 3 (poorly differentiated) categories. In routine practice, grade informs prognostic discussions and, together with other biomarkers, contributes to decisions about the broad classes of therapy a patient may be considered for.

The persistent problem with manual grading is reproducibility. Assigning tubule percentages, judging nuclear atypia, and especially counting mitoses are subjective tasks, and interobserver agreement — particularly for the large and clinically ambiguous grade 2 category — is only moderate. This has motivated efforts to build computational tools that could make grading more consistent, or at least provide a reproducible, automated prognostic readout from a standard hematoxylin and eosin (H&E) slide [1].

This is where deep-learning (DL) histologic grading enters. The key educational point of this article — and the central caveat for any clinician or trainee evaluating such tools — is a distinction that is easy to blur: a model can reproduce the prognostic information of manual grade without reproducing the individual case-level grade assignment. These are not the same thing, and conflating them can lead to serious misinterpretation.

What the method is and how it works

"predGrade" is a deep-learning model developed to perform Nottingham-like histologic grading of invasive breast carcinoma directly from whole-slide H&E images [1]. Rather than requiring a pathologist to score tubule formation, nuclei, and mitoses separately, the model is trained to predict a grade category from the digitized tissue image. Conceptually, it belongs to the family of image-based convolutional neural network approaches that learn morphologic features associated with tumor differentiation and outcome.

The intended output is a grade-like classification that stratifies patients by prognosis. The way such a model is validated matters enormously. There are two very different success criteria:

1. Concordance — does the model assign the same grade a pathologist would, case by case? This is measured by agreement statistics such as percent agreement and Cohen's kappa.
2. Prognostic stratification — do the model's grade groups separate patients into meaningfully different outcome trajectories, measured by hazard ratios (HRs) for survival or recurrence endpoints?

A tool can succeed on the second while only partially succeeding on the first. The predGrade evidence illustrates exactly this pattern, which is why it is framed as a prognosis-oriented tool rather than a replacement for manual grade assignment [1].

The evidence

The supporting study for predGrade is Sharma and colleagues (2024), published in Breast Cancer Research [1]. The model was developed and then evaluated on an external validation cohort of 1,262 cases, which is a reasonably sized independent test set and an important strength — external validation on cases not used in training is the minimum bar for claiming generalizability [1].

The headline finding is about prognosis. In that external cohort, the model's grading provided prognostic stratification with a hazard ratio comparable to that of the clinically assigned Nottingham histologic grade [1]. In practical terms, when patients were divided into grade groups by the algorithm, the separation in outcomes between those groups was of similar magnitude to the separation achieved when a pathologist assigns NHG manually. This is the core "SUPPORTED" claim: as a prognostic instrument, the AI model performed comparably to the established clinical grade [1].

The equally important finding — and the one that should temper enthusiasm — is that case-level agreement with manual grading was only modest [1]. In other words, if you look at any individual slide, the model and the pathologist will frequently disagree on whether it is grade 1, 2, or 3. The model does not simply "copy" the human grade. Instead, it appears to capture prognostically relevant morphology in its own way, producing groupings that happen to stratify outcomes similarly even though they do not map one-to-one onto human grade assignments [1].

This dissociation is the single most important concept for interpreting predGrade. It means two things simultaneously:

• The tool is prognosis-only: its value demonstrated in this study is risk stratification, not reproduction of the Nottingham grade label.
• The tool is not a concordance substitute: a clinician cannot assume that "the AI said grade 2" is interchangeable with "the pathologist would call this grade 2."

Because grade in clinical practice is frequently integrated with other variables to determine eligibility for broad therapy classes, this distinction has real downstream consequences. A prognostically equivalent but discordant grade could, for an individual patient, differ from the grade that pathologists and multidisciplinary teams are accustomed to using within those established decision frameworks.

Where it stands

Investigational. predGrade, as described in the cited study, is a research tool with external validation demonstrating prognostic performance comparable to clinical NHG [1]. The evidence presented does not establish it as a validated, regulatory-cleared diagnostic that can replace manual grading in routine reporting, and it should not be read that way. The evidence supports a specific, bounded claim — prognostic stratification on an external cohort of 1,262 cases — and no more [1].

Limitations and open questions. Several caveats follow directly from the evidence:

• The concordance gap. Modest case-level agreement with manual grade is not a minor footnote; it is central to how the tool should be understood. Prognostic equivalence at the group level does not guarantee correct classification at the individual level [1].
• Generalizability beyond the tested cohort. External validation on one cohort is encouraging but not definitive. Performance across different scanners, staining protocols, laboratories, and patient populations — and across clinically relevant subgroups such as receptor-defined subtypes — requires further study before broad deployment.
• Endpoint and follow-up. The strength of any survival-based claim depends on cohort maturity and endpoint definition; readers should interpret the reported HR within the design of the study that generated it [1].
• Role in practice. The evidence positions such a model as a potential adjunct for reproducible prognostic information, not as a stand-alone replacement for pathologist grading or for the composite workup used to determine therapy-class eligibility.

The evidence suggests that deep-learning grading can recover much of the prognostic signal embedded in histologic grade, in a reproducible and automated fashion, and that is a genuinely meaningful result for a field long troubled by grading variability [1]. But the same evidence cautions against treating the model's output as equivalent to a human grade on any given case. For now, predGrade is best understood as a validated-in-one-external-cohort, prognosis-oriented, investigational tool — promising as a reproducible risk stratifier, not yet a drop-in substitute for the Nottingham grade a patient and clinician will see on a pathology report [1].

---

References

1. Sharma Y, et al. Deep learning–based Nottingham-like histologic grading of breast carcinoma (predGrade): prognostic stratification comparable to clinical Nottingham histologic grade with only modest case-level agreement, external validation cohort n = 1,262. Breast Cancer Research. 2024. PMID 38287342; DOI 10.1186/s13058-024-01770-4.