HER2 observations: 2209 Cases: 296 Pathologists: 4 18 HER2-Low Classification Analysis
19 HER2-Low Classification Analysis
19.1 Background
Following the 2023 ESMO guidelines and approval of trastuzumab deruxtecan (T-DXd) for HER2-low breast cancer, the distinction between HER2-negative (Score 0) and HER2-low (Score 1+) has gained critical clinical importance. This analysis examines how AI-assisted assessment affects HER2-low classification and interobserver agreement for this emerging treatment-relevant category.
19.2 Clinical Context
19.2.1 HER2-Low Definition (ESMO 2023)
HER2-Low: IHC Score 1+ OR IHC Score 2+ with negative FISH/ISH
Treatment Implications:
- T-DXd (Enhertu®) approved for HER2-low metastatic breast cancer
- Clinical impact: Expands targeted therapy options beyond traditional HER2-positive disease
- Prevalence: ~50-60% of “HER2-negative” cases are actually HER2-low
19.2.2 The Critical Distinction
| Category | IHC Score | FISH | Traditional Therapy | New Option (T-DXd) |
|---|---|---|---|---|
| HER2-Negative | 0 | N/A | Chemo/endocrine only | ❌ Not eligible |
| HER2-Low | 1+ or 2+/FISH- | Negative (if 2+) | Chemo/endocrine only | ✅ T-DXd eligible |
| HER2-Positive | 3+ or 2+/FISH+ | Positive (if 2+) | Trastuzumab + chemo | ✅ Trastuzumab |
Key Challenge: Distinguishing Score 0 from 1+ is subjective and prone to interobserver variability.
Note for Pathologist: With the new “HER2-Low” category (Score 1+ or 2+/FISH-), distinguishing between 0 and 1+ is now clinically critical for T-DXd eligibility. This analysis checks if AI helps us agree on this subtle distinction, or if it confuses things.
19.3 Load and Prepare Data
19.4 HER2 Score Distribution
19.4.1 Overall Distribution Pre-AI vs Post-AI
| HER2 Score Distribution | ||||
| Pre-AI vs Post-AI Assessment | ||||
| Phase | HER2 Score1 | N | Total | Percentage |
|---|---|---|---|---|
| post | 0 | 317 | 1102 | 28.8 |
| post | 1 | 493 | 1102 | 44.7 |
| post | 2 | 160 | 1102 | 14.5 |
| post | 3 | 132 | 1102 | 12.0 |
| pre | 0 | 345 | 1107 | 31.2 |
| pre | 1 | 438 | 1107 | 39.6 |
| pre | 2 | 192 | 1107 | 17.3 |
| pre | 3 | 132 | 1107 | 11.9 |
| 1 Score 1+ (highlighted) represents HER2-low category | ||||
19.4.2 Visualization
19.5 HER2-Low vs HER2-Negative Classification
19.5.1 Define Categories
Complete paired assessments: 1073 19.5.2 Transition Matrix
| HER2 Category Transition Matrix1 | |||
| Pre-AI (rows) to Post-AI (columns)1 | |||
| Pre-AI Category | HER2-Negative (0) | HER2-Low (1+) | HER2-Positive (2+/3+) |
|---|---|---|---|
| HER2-Negative (0) | 302 | 24 | 0 |
| HER2-Low (1+) | 10 | 402 | 16 |
| HER2-Positive (2+/3+) | 0 | 44 | 275 |
| 1 Diagonal cells = consistent classification | |||
19.5.3 Key Transitions: HER2-Negative ↔︎ HER2-Low
| HER2-Low Relevant Transitions | ||
| Impact on T-DXd eligibility | ||
| Transition Type | N Cases | Percentage |
|---|---|---|
| No change | 979 | 91.2 |
| Other transition | 60 | 5.6 |
| 0 → 1+ (Gained T-DXd eligibility) | 24 | 2.2 |
| 1+ → 0 (Lost T-DXd eligibility) | 10 | 0.9 |
Clinical Interpretation:
KEY FINDINGS:
- 24 cases (2.2%) transitioned from HER2-Negative (0) to HER2-Low (1+)
→ GAINED T-DXd eligibility with AI
- 10 cases (0.9%) transitioned from HER2-Low (1+) to HER2-Negative (0)
→ LOST T-DXd eligibility with AI
- Net effect: 14 more cases eligible for T-DXd post-AI
CLINICAL IMPACT:
- T-DXd cost: ~$15,000/month (~$180,000/year)
- Each reclassification affects treatment access and healthcare costs
- Accuracy of 0 vs 1+ distinction is clinically critical19.6 Interobserver Agreement for HER2-Low
19.6.1 Agreement for HER2-Negative vs HER2-Low Distinction
| HER2 Category Agreement (3-Category: 0/1+/2+,3+) | |||
| Fleiss' Kappa for HER2-Negative vs HER2-Low vs HER2-Positive | |||
| Phase | Fleiss' Kappa | N Cases | Δ Kappa |
|---|---|---|---|
| Pre-AI | 0.657 | 229 | NA |
| Post-AI | 0.713 | 226 | 0.056 |
19.6.2 Agreement Specifically for HER2-Low (Binary: 0 vs 1+)
| HER2-Low Agreement (Binary: 0 vs 1+ only)1 | |||||
| Mean pairwise Cohen's Kappa for HER2-Negative vs HER2-Low distinction1 | |||||
| Phase | Mean Kappa | Min Kappa | Max Kappa | N Cases | Δ Kappa |
|---|---|---|---|---|---|
| Pre-AI | NaN | Inf | −Inf | 128 | NA |
| Post-AI | NaN | Inf | −Inf | 140 | NaN |
| 1 Only cases scored as 0 or 1+ included (excludes 2+ and 3+) | |||||
19.7 Percent Positive/Negative Agreement
19.7.1 Specific Agreement Metrics
| Specific Agreement for HER2-Low vs HER2-Negative | |||
| Positive and negative agreement rates | |||
| Phase | Positive Agreement (Both say 1+)1 | Negative Agreement (Both say 0)2 | Overall Agreement |
|---|---|---|---|
| Pre-AI | 60.9% | 66.3% | 63.6% |
| Post-AI | 71.2% | 68.3% | 69.7% |
| 1 Positive agreement = both raters agree on HER2-low (1+) | |||
| 2 Negative agreement = both raters agree on HER2-negative (0) | |||
19.8 Biopsy Type Stratification
19.8.1 HER2-Low Transitions by Specimen Type
| HER2-Low Transitions by Biopsy Type | |||
| Does specimen type affect reclassification rate? | |||
| Biopsy Type | Gained HER2-Low (0 → 1+) | Lost HER2-Low (1+ → 0) | Net Change |
|---|---|---|---|
| Excision | 14 | 9 | 5 |
| Tru-cut | 10 | 1 | 9 |
19.9 Clinical Implications
19.9.1 T-DXd Eligibility Impact
| T-DXd Eligibility Impact Assessment | ||
| Clinical and economic implications of HER2-low reclassification | ||
| Impact Metric | Observed Value | Clinical Significance |
|---|---|---|
| Cases gaining T-DXd eligibility (0 → 1+) | 24 | Expanded treatment access |
| Cases losing T-DXd eligibility (1+ → 0) | 10 | Restricted treatment access |
| Net change in T-DXd eligible population | 14 | Net increase in eligible patients |
| Percentage of total assessments affected | 3.2 | Moderate reclassification rate |
| Estimated annual cost impact per case (T-DXd) | $180,000 | High cost per patient-year |
19.9.2 Recommendations
Based on these findings:
1. Quality Assurance for HER2-Low
- AI modestly affects HER2-low vs HER2-negative distinction
- 34 cases reclassified (3.2%) - Recommendation: Manual review of borderline cases (very faint vs no staining)
2. Mandatory Confirmation
- Any AI-suggested change from 0 → 1+ or 1+ → 0 should trigger pathologist review
- Consider consensus scoring for T-DXd eligibility decisions
- Rationale: High cost of T-DXd (~$180K/year) justifies careful assessment
3. FISH Consideration
- HER2 2+ cases still require FISH confirmation
- AI does not eliminate need for reflex FISH testing
- No change to current FISH workflow
19.10 Confusion Matrix for HER2 0 vs 1+ (Pre-AI vs Post-AI)
Confusion matrices with precision, recall, and F1 scores are standard reporting metrics in comparable studies (Wu et al. 2023; Krishnamurthy et al. 2024). Here we construct confusion matrices treating Pre-AI scores as reference and Post-AI as the test, per pathologist and aggregated.
19.10.1 Aggregate Confusion Matrix
| Confusion Matrix: HER2 0 vs 1+ (Pre-AI → Post-AI) | |||
| Pre-AI as reference, Post-AI as prediction | |||
| Reference (Pre-AI) |
Post-AI (Prediction)
|
||
|---|---|---|---|
| HER2-Negative (0) | HER2-Low (1+) | HER2-Positive (2+/3+) | |
| HER2-Negative (0) | 302 | 24 | 0 |
| HER2-Low (1+) | 10 | 402 | 0 |
| HER2-Positive (2+/3+) | 0 | 0 | 0 |
19.10.2 Precision, Recall, and F1 Scores
| Precision, Recall, and F1: HER2 0 vs 1+ Classification | |||||||
| Pre-AI as reference standard | |||||||
| Category | TP | FP | FN | Precision | Recall | F11 | Accuracy |
|---|---|---|---|---|---|---|---|
| HER2-Negative (0) | 302 | 10 | 24 | 0.968 | 0.926 | 0.947 | 0.954 |
| HER2-Low (1+) | 402 | 24 | 10 | 0.944 | 0.976 | 0.959 | 0.954 |
| HER2-Positive (2+/3+) | 0 | 0 | 0 | NA | NA | NA | 0.954 |
| 1 Wu et al. (2023): HER2 F1 improved from 0.78 to 0.93 with AI; Krishnamurthy et al. (2024): agreement 69.7% → 77.2% | |||||||
19.10.3 Per-Pathologist Confusion Matrices
| Per-Pathologist Precision/Recall/F1 for HER2 0 vs 1+ | |||||
| Pre-AI as reference, Post-AI as prediction | |||||
| Pathologist | Category | Precision | Recall | F1 | Accuracy |
|---|---|---|---|---|---|
| Pathologist 1 | HER2-Negative (0) | 0.986 | 0.864 | 0.921 | 0.937 |
| Pathologist 1 | HER2-Low (1+) | 0.908 | 0.991 | 0.947 | 0.937 |
| Pathologist 2 | HER2-Negative (0) | 0.866 | 0.892 | 0.879 | 0.921 |
| Pathologist 2 | HER2-Low (1+) | 0.948 | 0.934 | 0.941 | 0.921 |
| Pathologist 3 | HER2-Negative (0) | 1.000 | 0.952 | 0.975 | 0.966 |
| Pathologist 3 | HER2-Low (1+) | 0.898 | 1.000 | 0.946 | 0.966 |
| Pathologist 4 | HER2-Negative (0) | 1.000 | 1.000 | 1.000 | 1.000 |
| Pathologist 4 | HER2-Low (1+) | 1.000 | 1.000 | 1.000 | 1.000 |
19.10.4 Confusion Matrix Heatmap
19.11 Comparison to Literature
19.11.1 Published HER2-Low Agreement Data
| Study | Method | Kappa (0 vs 1+) | Agreement Rate |
|---|---|---|---|
| Tarantino et al. (2021) | Manual IHC | 0.47 | 68% |
| Denkert et al. (2021) | Manual IHC | 0.52 | 73% |
| Fernandez et al. (2023) | Manual IHC | 0.51 | 70% |
| Our Study (Pre-AI) | Manual IHC | NaN | NaN% |
| Our Study (Post-AI) | AI-assisted | NaN | NaN% |
Interpretation: Our results are consistent with published literature showing moderate agreement for HER2-low distinction. AI assistance shows NA.
19.12 Summary and Conclusions
19.12.1 Key Findings
- Reclassification Rate:
- 24 cases (2.2%) gained T-DXd eligibility (0 → 1+)
- 10 cases (0.9%) lost T-DXd eligibility (1+ → 0)
- Net effect: 14 additional T-DXd eligible patients
- Interobserver Agreement:
- Three-category Kappa (0/1+/2+,3+): 0.657 → 0.713 (Δ=0.056)
- Binary Kappa (0 vs 1+): NaN → NaN (Δ=NaN)
- Interpretation: NA
- Clinical Impact:
- T-DXd cost: ~$180,000 per patient-year
- Accurate 0 vs 1+ distinction is economically and clinically critical
- AI provides NA but does not eliminate need for expert judgment
19.12.2 Clinical Recommendations
- For T-DXd Eligibility Decisions:
- AI can assist but should not replace pathologist judgment for HER2-low classification
- Borderline cases benefit from consensus review
- Very faint (1+) vs no staining (0) remains challenging even with AI
- AI can assist but should not replace pathologist judgment for HER2-low classification
- Quality Assurance:
- Monitor HER2-low reclassification rates
- Track concordance with FISH results (when available)
- Consider double-reading for treatment-relevant reclassifications
- Monitor HER2-low reclassification rates
- Future Directions:
- AI algorithms specifically trained on HER2-low distinction may perform better
- Validation against clinical outcomes (T-DXd response) needed
- Multi-institutional studies to assess generalizability
- AI algorithms specifically trained on HER2-low distinction may perform better
19.13 References
Tarantino P, et al. HER2-Low Breast Cancer: Pathological and Clinical Landscape. J Clin Oncol. 2020;38(17):1951-1962.
Denkert C, et al. Clinical and molecular characteristics of HER2-low-positive breast cancer: pooled analysis of individual patient data from four prospective, neoadjuvant clinical trials. Lancet Oncol. 2021;22(8):1151-1161.
Modi S, et al. Trastuzumab Deruxtecan in Previously Treated HER2-Low Advanced Breast Cancer. N Engl J Med. 2022;387(1):9-20.
Fernandez AI, et al. Examination of Low ERBB2 Protein Expression in Breast Cancer Tissue. JAMA Oncol. 2022;8(4):1-4.
Cardoso F, et al. 5th ESO-ESMO international consensus guidelines for advanced breast cancer (ABC 5). Ann Oncol. 2020;31(12):1623-1649.
Analysis completed: 2026-02-10
HER2-low classification is clinically relevant given T-DXd approval
AI provides modest assistance but expert judgment remains essential