13  Subgroup Analysis

13.1 Objective

Analyze AI impact and agreement patterns within specific subgroups defined by tumor characteristics.

Note for Pathologist: Breast cancer is not one disease. We break down the performance by molecular subtype (e.g., Triple Negative vs Luminal A) and by “borderline” status (e.g., Ki67 near 20-30%). This ensures we don’t claim “Good Accuracy” overall while hiding poor performance in the most critical, difficult subgroups.

13.2 Setup

13.3 Load Data

13.4 Agreement by Molecular Subtype

Does inter-observer agreement vary by molecular subtype?

Inter-Observer Agreement by Molecular Subtype (Pre-AI)
ICC Values for Continuous Markers
Molecular Subtype N Cases ER ICC PR ICC Ki67 ICC
HER2 Positive 324 0.971 0.936 0.933
Hormone Weak Positive 189 0.415 0.993 0.851
Luminal A 390 0.655 0.818 0.670
Luminal B 146 0.735 0.953 0.767
Triple Negative 135 NaN NaN 0.949
Inter-Observer Agreement by Molecular Subtype (Post-AI)
ICC Values for Continuous Markers
Molecular Subtype N Cases ER ICC PR ICC Ki67 ICC
HER2 Positive 292 0.984 0.975 0.921
Hormone Weak Positive 173 0.784 0.991 0.890
Luminal A 310 0.424 0.946 0.647
Luminal B 268 0.969 0.976 0.783
Triple Negative 141 NaN NaN 0.939

Note for Pathologist: ICC values are shown separately for each molecular subtype. Low ICC (below 0.75) suggests poor agreement within that subgroup. If ICC improves from Pre-AI to Post-AI for a given subtype, it means AI helped pathologists agree more on cases of that type. Subtypes with very few cases (small N) may produce unreliable ICC estimates.

13.5 AI Impact by HER2 Status

Compare AI influence in HER2 positive vs negative cases.

AI Impact by HER2 Status
Mean absolute changes in marker values
HER2 Status N Cases ER MAE PR MAE Ki67 MAE HER2 Changes (%)
HER2 Negative 783 3.07 4.54 6.76 6.63
HER2 Positive/Equivocal 324 3.09 2.89 7.29 15.36

13.6 Agreement in Hormone Receptor Positive vs Negative

Agreement by Hormone Receptor Status
ICC comparison in HR+ vs HR- cases
HR Status N Cases ER ICC PR ICC Ki67 ICC Timing
HR Negative 193 NaN NaN 0.950 Pre-AI
HR Positive 988 0.815 0.943 0.900 Pre-AI
HR Negative 180 NaN NaN 0.927 Post-AI
HR Positive 994 0.917 0.969 0.911 Post-AI

13.7 Ki67 Proliferation Categories

Analyze agreement in high vs low proliferation cases.

Agreement by Ki67 Proliferation Category
Pre-AI vs Post-AI ICC values
Ki67 Category N Cases ER Pre ER Post PR Pre PR Post Ki67 Pre Ki67 Post
High (≥30%) 399 0.978 0.989 0.943 0.964 0.881 0.868
Intermediate (20-30%) 193 0.923 0.954 0.948 0.893 0.745 0.523
Low (<20%) 583 0.913 0.947 0.963 0.979 0.707 0.700
AI Impact by Ki67 Category
Mean absolute changes and subtype reclassification
Ki67 Category N Cases ER Change PR Change Ki67 Change Subtype Changed (%)
High (≥30%) 399 3.00 2.63 7.62 9.77
Intermediate (20-30%) 193 3.18 4.89 8.84 44.56
Low (<20%) 583 3.03 4.84 5.72 13.55

13.8 Borderline Cases Analysis

Focus on cases near clinical thresholds where AI impact may be most critical.

Note for Pathologist: “Borderline” cases sit near critical clinical cutoffs (e.g., ER 0-10%, Ki67 25-35%, HER2 Score 2). These are the cases where small changes in scoring can change the treatment plan entirely. If AI improves agreement in borderline cases specifically, it provides the greatest clinical value.

Borderline case statistics:
ER borderline (0-10%): 228 
PR borderline (0-10%): 634 
Ki67 borderline (25-35%): 193 
HER2 borderline (Score 2): 192 
Any borderline: 799
AI Impact: Borderline vs Non-Borderline Cases
Cases near clinical thresholds
N Cases ER Change PR Change Ki67 Change HER2 Changed (%) Subtype Changed (%) Case Type
341 3.38 7.90 7.09 6.42 19.06 Non-Borderline
799 2.95 2.30 6.83 10.54 16.65 Borderline
44 2.98 7.39 6.13 0.00 13.64 NA
Agreement in Borderline Cases
ICC values comparing borderline vs non-borderline
N Cases ER Pre ER Post PR Pre PR Post Ki67 Pre Ki67 Post Case Type
341 0.410 0.604 0.804 0.864 0.933 0.918 Non-Borderline
799 0.973 0.987 0.929 0.958 0.944 0.946 Borderline
44 NA NA NA NA NA NA NA

13.9 Triple Negative Cases

Special focus on triple negative breast cancer (TNBC) cases. TNBC is characterized by aggressive behavior and lack of targeted therapies, making accurate biomarker assessment and identification of potential responders to chemotherapy (e.g., via Ki67) critical. AI approaches have been highlighted as promising tools for refining prognosis and prediction in this subgroup (Corredor et al. 2023).

Triple Negative cases:
Pre-AI: 135 
Post-AI: 141 
Any TNBC designation: 143 
Cases reclassified to/from TNBC: 10
TNBC Reclassifications
Cases that changed to or from Triple Negative status
case_id pathologist molecular_subtype_pre molecular_subtype_post er_pre er_post pr_pre pr_post her2_pre her2_post
21524-25 Pathologist 2 Triple Negative Hormone Weak Positive 0.0 1.0 0.0 0.0 1 1
18413-25 Pathologist 2 HER2 Positive Triple Negative 0.0 0.0 0.0 0.0 2 1
14982-25 Pathologist 2 HER2 Positive Triple Negative 0.0 0.0 0.0 0.0 2 1
9229-25 Pathologist 2 Luminal A Triple Negative 90.0 NA 10.0 NA 1 NA
6476-25 Pathologist 2 HER2 Positive Triple Negative 0.0 0.0 0.0 0.0 2 1
25564-25 Pathologist 1 HER2 Positive Triple Negative 0.0 0.0 0.0 0.0 2 1
11985-25 Pathologist 1 Triple Negative Hormone Weak Positive 0.0 2.0 0.0 0.0 0 0
16127-25 Pathologist 3 Luminal A Triple Negative 80.0 NA 60.0 NA 1 NA
14982-25 Pathologist 3 HER2 Positive Triple Negative 0.0 0.0 0.0 0.0 2 NA
11354-25 Pathologist 3 Luminal A Triple Negative 90.0 NA 20.0 NA 1 NA 

Agreement in TNBC cases:
# A tibble: 1 × 7
  n_cases er_icc_pre er_icc_post pr_icc_pre pr_icc_post ki67_icc_pre
    <int>      <dbl>       <dbl>      <dbl>       <dbl>        <dbl>
1     143        NaN    -0.00868        NaN         NaN        0.949
# ℹ 1 more variable: ki67_icc_post <dbl>

13.10 Luminal A vs B Differentiation

Focus on cases where Ki67 determines Luminal subtype.

Luminal A/B reclassifications: 85
Luminal A ↔︎ B Transitions
Cases that switched between Luminal subtypes
case_id pathologist molecular_subtype_pre molecular_subtype_post ki67_pre ki67_post er_pre er_post pr_pre pr_post
33681-25 Pathologist 2 Luminal B Luminal A 35.0 25.0 50.0 50.0 30.0 30.0
32252-25 Pathologist 2 Luminal A Luminal B 25.0 33.0 100.0 100.0 100.0 90.0
31276-25 Pathologist 2 Luminal A Luminal B 29.0 30.0 100.0 100.0 70.0 55.0
31467-25 Pathologist 2 Luminal A Luminal B 20.0 30.0 95.0 95.0 50.0 60.0
30689-25 Pathologist 2 Luminal A Luminal B 23.0 35.0 95.0 100.0 80.0 85.0
24570-25 Pathologist 2 Luminal A Luminal B 23.0 32.0 90.0 89.0 60.0 41.0
21963-25 Pathologist 2 Luminal A Luminal B 27.0 37.0 95.0 95.0 95.0 95.0
20823-25 Pathologist 2 Luminal A Luminal B 21.0 30.0 80.0 77.0 60.0 54.0
18676-25 Pathologist 2 Luminal A Luminal B 27.0 40.0 90.0 93.0 30.0 17.0
15305-25 Pathologist 2 Luminal A Luminal B 28.0 44.0 90.0 86.0 70.0 69.0
14815-25 Pathologist 2 Luminal A Luminal B 28.0 30.0 100.0 95.0 50.0 45.0
14262-25 Pathologist 2 Luminal A Luminal B 25.0 35.0 95.0 95.0 30.0 20.0
11354-25 Pathologist 2 Luminal A Luminal B 25.0 31.0 95.0 95.0 40.0 25.0
9346-25 Pathologist 2 Luminal A Luminal B 16.0 30.0 90.0 92.0 40.0 50.0
6444-25 Pathologist 2 Luminal A Luminal B 25.0 50.0 90.0 90.0 90.0 80.0
33988-25 Pathologist 1 Luminal A Luminal B 6.0 32.0 95.0 97.0 95.0 97.0
33813-25 Pathologist 1 Luminal B Luminal A 32.0 21.0 95.0 99.0 100.0 99.0
32396-25 Pathologist 1 Luminal A Luminal B 20.0 30.0 95.0 98.0 50.0 50.0
31448-25 Pathologist 1 Luminal A Luminal B 26.0 37.0 90.0 91.0 50.0 47.0
30689-25 Pathologist 1 Luminal A Luminal B 8.0 35.0 95.0 89.0 40.0 59.0
29103-25 Pathologist 1 Luminal A Luminal B 24.0 69.0 100.0 96.0 80.0 63.0
28463-25 Pathologist 1 Luminal A Luminal B 22.0 45.0 70.0 45.0 30.0 43.0
26562-25 Pathologist 1 Luminal A Luminal B 12.0 45.0 95.0 95.0 70.0 55.0
24570-25 Pathologist 1 Luminal A Luminal B 23.0 32.0 90.0 90.0 50.0 40.0
21963-25 Pathologist 1 Luminal A Luminal B 15.0 34.0 85.0 90.0 90.0 90.0
22030-25 Pathologist 1 Luminal A Luminal B 16.0 35.0 95.0 95.0 95.0 80.0
20019-25 Pathologist 1 Luminal A Luminal B 16.0 30.0 100.0 95.0 95.0 79.0
18267-25 Pathologist 1 Luminal A Luminal B 11.0 31.0 95.0 92.0 60.0 42.0
16396-25 Pathologist 1 Luminal A Luminal B 24.0 45.0 95.0 98.0 80.0 70.0
16497-25 Pathologist 1 Luminal A Luminal B 15.0 30.0 90.0 93.0 60.0 70.0
14815-25 Pathologist 1 Luminal A Luminal B 28.0 30.0 95.0 95.0 70.0 50.0
14987-25 Pathologist 1 Luminal A Luminal B 26.0 44.0 80.0 92.0 15.0 9.0
14262-25 Pathologist 1 Luminal A Luminal B 25.0 36.0 90.0 95.0 30.0 19.0
14074-25 Pathologist 1 Luminal A Luminal B 12.0 37.0 90.0 92.0 90.0 86.0
13388-25 Pathologist 1 Luminal A Luminal B 15.0 32.0 95.0 96.0 95.0 96.0
13472-25 Pathologist 1 Luminal A Luminal B 24.0 30.0 70.0 94.0 50.0 58.0
11625-25 Pathologist 1 Luminal A Luminal B 12.0 40.0 95.0 82.0 95.0 70.0
10928-25 Pathologist 1 Luminal A Luminal B 10.0 35.0 95.0 86.0 70.0 48.0
9753-25 Pathologist 1 Luminal A Luminal B 11.0 30.0 90.0 94.0 80.0 71.0
7121-25 Pathologist 1 Luminal A Luminal B 27.0 38.0 90.0 95.0 30.0 17.0
6288-25 Pathologist 1 Luminal A Luminal B 13.0 30.0 95.0 90.0 95.0 73.0
33836-25 Pathologist 4 Luminal A Luminal B 29.0 40.0 95.0 95.0 90.0 90.0
33988-25 Pathologist 4 Luminal A Luminal B 28.0 31.0 95.0 95.0 90.0 90.0
32396-25 Pathologist 4 Luminal A Luminal B 20.0 40.0 95.0 95.0 80.0 70.0
32252-25 Pathologist 4 Luminal A Luminal B 27.0 33.0 95.0 95.0 95.0 90.0
31276-25 Pathologist 4 Luminal A Luminal B 25.0 34.0 95.0 95.0 50.0 50.0
30689-25 Pathologist 4 Luminal A Luminal B 18.0 34.0 95.0 95.0 90.0 90.0
28463-25 Pathologist 4 Luminal A Luminal B 22.0 45.0 95.0 80.0 40.0 50.0
24570-25 Pathologist 4 Luminal A Luminal B 23.0 32.0 95.0 90.0 40.0 40.0
22438-25 Pathologist 4 Luminal A Luminal B 28.0 30.0 95.0 95.0 90.0 95.0
20019-25 Pathologist 4 Luminal A Luminal B 16.0 30.0 95.0 95.0 95.0 90.0
18217-25 Pathologist 4 Luminal A Luminal B 22.0 36.0 95.0 95.0 70.0 70.0
17530-25 Pathologist 4 Luminal A Luminal B 28.0 35.0 80.0 80.0 60.0 70.0
16396-25 Pathologist 4 Luminal A Luminal B 24.0 55.0 95.0 95.0 70.0 70.0
16497-25 Pathologist 4 Luminal A Luminal B 15.0 30.0 90.0 90.0 70.0 70.0
15305-25 Pathologist 4 Luminal A Luminal B 22.0 42.0 80.0 80.0 60.0 70.0
14815-25 Pathologist 4 Luminal A Luminal B 28.0 30.0 95.0 95.0 60.0 60.0
14262-25 Pathologist 4 Luminal A Luminal B 26.0 35.0 95.0 95.0 10.0 20.0
13388-25 Pathologist 4 Luminal A Luminal B 20.0 32.0 95.0 95.0 95.0 95.0
11625-25 Pathologist 4 Luminal A Luminal B 18.0 39.0 90.0 80.0 80.0 70.0
11354-25 Pathologist 4 Luminal A Luminal B 21.0 30.0 95.0 90.0 30.0 20.0
9753-25 Pathologist 4 Luminal A Luminal B 27.0 30.0 80.0 90.0 40.0 60.0
7121-25 Pathologist 4 Luminal A Luminal B 28.0 34.0 95.0 95.0 40.0 30.0
33988-25 Pathologist 3 Luminal A Luminal B 6.0 30.0 90.0 90.0 90.0 90.0
33813-25 Pathologist 3 Luminal B Luminal A 30.0 21.0 90.0 90.0 90.0 90.0
32396-25 Pathologist 3 Luminal A Luminal B 20.0 40.0 90.0 90.0 60.0 60.0
31448-25 Pathologist 3 Luminal A Luminal B 26.0 40.0 80.0 80.0 60.0 40.0
31276-25 Pathologist 3 Luminal A Luminal B 25.0 30.0 90.0 90.0 50.0 50.0
31467-25 Pathologist 3 Luminal A Luminal B 7.0 30.0 90.0 90.0 50.0 60.0
30689-25 Pathologist 3 Luminal A Luminal B 23.0 33.0 90.0 90.0 60.0 60.0
28463-25 Pathologist 3 Luminal A Luminal B 22.0 40.0 90.0 70.0 50.0 40.0
26562-25 Pathologist 3 Luminal A Luminal B 12.0 44.0 90.0 80.0 40.0 50.0
24570-25 Pathologist 3 Luminal A Luminal B 10.0 32.0 90.0 90.0 20.0 30.0
21963-25 Pathologist 3 Luminal A Luminal B 26.0 30.0 80.0 90.0 80.0 80.0
22030-25 Pathologist 3 Luminal A Luminal B 16.0 30.0 80.0 90.0 70.0 80.0
20019-25 Pathologist 3 Luminal A Luminal B 15.0 30.0 80.0 90.0 80.0 80.0
16396-25 Pathologist 3 Luminal A Luminal B 23.0 30.0 90.0 95.0 80.0 70.0
14262-25 Pathologist 3 Luminal A Luminal B 25.0 30.0 90.0 85.0 20.0 20.0
14095-25 Pathologist 3 Luminal A Luminal B 10.0 30.0 90.0 90.0 30.0 40.0
13388-25 Pathologist 3 Luminal A Luminal B 25.0 31.0 90.0 90.0 90.0 90.0
13472-25 Pathologist 3 Luminal A Luminal B 25.0 30.0 90.0 90.0 60.0 60.0
10676-25 Pathologist 3 Luminal A Luminal B 23.0 33.0 80.0 90.0 80.0 80.0
9753-25 Pathologist 3 Luminal A Luminal B 20.0 30.0 80.0 80.0 80.0 70.0
7121-25 Pathologist 3 Luminal A Luminal B 27.0 35.0 80.0 90.0 10.0 15.0
6288-25 Pathologist 3 Luminal A Luminal B 15.0 30.0 70.0 60.0 50.0 50.0
Ki67 Values in Luminal A/B Transitions
mean_ki67_pre mean_ki67_post median_ki67_pre median_ki67_post
21.1 34.5 23.0 32.0

13.11 Conclusion

13.11.1 Key Findings from Subgroup Analysis

  1. Molecular Subtype Dependency:
    • Agreement metrics vary substantially across molecular subtypes
    • AI impact differs by subtype, with certain subtypes showing more changes
  2. HER2 Status Impact:
    • HER2+ cases may show different agreement patterns than HER2- cases
    • AI influence varies by HER2 status
  3. Hormone Receptor Status:
    • HR+ and HR- cases show different baseline agreement levels
    • AI may have differential impact based on HR status
  4. Proliferation Categories:
    • Ki67 categories affect both agreement and AI influence
    • High proliferation cases may be more challenging
  5. Borderline Cases:
    • Cases near clinical thresholds warrant special attention
    • AI impact is often strongest in borderline cases
    • These cases have the highest clinical impact
  6. TNBC Identification:
    • Reclassifications to/from TNBC are critically important
    • Requires careful review due to treatment implications
  7. Luminal Subtype Differentiation:
    • Ki67 cutoff at 30% drives Luminal A vs B classification
    • AI-influenced Ki67 changes can significantly affect treatment decisions

13.11.2 Clinical Recommendations

  • Subgroup-specific quality metrics should be monitored
  • Borderline and reclassified cases require additional review
  • AI performance should be validated separately for each subgroup
  • Training data for AI should include adequate representation of all subgroups