The Combined Analysis of Pathology and Artificial Intelligence (AI; CAPAI) model effectively stratified patients with colon cancer into distinct prognostic groups, identifying nearly half as low-risk, with “favorable” cancer-specific survival outcomes in the absence of adjuvant chemotherapy, according to presenting author Marie-Christine E. Bakker, MS, of the University Medical Center Utrecht, the Netherlands. These findings from an interim report of a Dutch nationwide validation study were presented at the European Society for Medical Oncology (ESMO) Congress 2025.1
“Prospective use of this readily available and affordable stratification tool could help avoid adjuvant chemotherapy in low-risk patients and intensify treatment in high-risk patients,” commented Ms. Bakker and colleagues.
Marie-Christine E. Bakker, MS
The recommended management of high-risk stage II and III colon cancer consists of surgical resection followed by adjuvant chemotherapy.2 Nevertheless, Ms. Bakker noted that approximately half of these patients achieve cure with surgery alone and are therefore overtreated, which, she explained, “exposes them to unnecessary toxicity risks and [imposes] costs on society.” She thus emphasized the need for optimization of risk stratification.
“In the development and validation study, CAPAI yielded a hazard ratio [for cancer-specific survival] of 10.71 [CAPAI-defined high- vs low-risk patients treated with adjuvant chemotherapy],” Ms. Bakker said.3 “However, it is important to study the performance of CAPAI in other cohorts as well, with a special interest in a cohort of patients that did not receive adjuvant chemotherapy…,” providing rationale for the present analysis.
Study Details
The investigators used the Netherlands Cancer Registry to identify eligible patients—those with high-risk stage II or III colon cancer who underwent R0 resection between 2015 and 2019 and did not receive (neo)adjuvant chemotherapy despite being fit (World Health Organization 0–1; American Society of Anesthesiologists 1–2; and age < 70 years). Data on baseline characteristics and outcomes were obtained from the registry. Corresponding diagnostic hematoxylin-and-eosin slides were collected via the Nationwide Automated Pathology Archive, centrally scanned, and analyzed using the DoMore-v1 deep learning biomarker platform; this platform assesses the entire tumor microenvironment and links morphologic patterns to survival outcomes.
This analysis included 453 patients from 65 hospitals—a “substantial cohort” because of the restrictive use of adjuvant chemotherapy in the Netherlands, Ms. Bakker noted. Among this population, 88 had stage II high-risk disease, 272 had stage III low-risk disease, and 93 had stage III high-risk disease. A total of 251, 55, and 147 patients were classified as having DoMore-v1–defined good, uncertain, and poor prognosis, respectively.
“After integrating these AI categories with pathologic staging markers [pT/pN stage and lymph node count], CAPAI eventually classified 215 as low risk, 156 as intermediate risk, and 82 as high risk,” Ms. Bakker remarked, “so almost half of the patients were classified as low risk.”
Cancer-Specific Survival
The 3-year cancer-specific survival (primary endpoint; defined as death with evidence of recurrence) rate was 85.9% (95% confidence interval [CI] = 82.6%–89.2%). In the CAPAI-defined low-, intermediate-, and high-risk groups, the rates were 93.7% (95% CI = 90.5%–97.1%), 87.5% (95% CI = 82.3%–93.1%), and 60.4% (95% CI = 50.4%–72.5%), respectively. Univariable Cox regression showed a hazard ratio for intermediate- vs low-risk patients of 2.1 (95% CI = 1.0–4.3; P = .039); for those with high- vs low-risk disease, the hazard ratio was 7.9 (95% CI = 4.1–15.2; P < .001). According to Ms. Bakker, these results “highlight the favorable survival of the low-risk group…, despite not having received adjuvant chemotherapy.”
Clinical Implications
“Multimodal risk assessment with combined pathology and AI allows identification of a clinically relevant low-risk group of 47.5% of locally advanced colon cancer, with a 3-year cancer-specific survival of 93.7% without adjuvant chemotherapy,” Ms. Bakker concluded.
According to Ms. Bakker, the clinical implications include personalizing treatment for locally advanced colon cancer and safely deescalating adjuvant chemotherapy for low-risk patients. “However,” she noted, “before clinical implementation, prospective validation is needed.”
A Dutch nationwide validation study of CAPAI in patients treated with and without adjuvant chemotherapy is ongoing. Additionally, an intervention study is being prepared, combining CAPAI with additional biomarkers (eg, circulating tumor DNA).
DISCLOSURE: Ms. Bakker reported financial interests with, institutional affiliations with, and research funding from Health~Holland. For full disclosures of the study authors, visit cslidectimeetingtech.com.
REFERENCES
1. Bakker MCE, Franken IA, Laclé MM, et al: Prognostic value of the combined analysis of pathologists and artificial intelligence (CAPAI) in high-risk stage II-III colon cancer treated without chemotherapy: Interim report from a nationwide validation. ESMO Congress 2025. Abstract 726O. Presented October 20, 2025.
2. Argilés G, Tabernero J, Labianca R, et al: Localised colon cancer: ESMO Clinical Practice Guidelines for diagnosis, treatment and follow-up. Ann Oncol 31:1291-1305, 2020.
3. Kleppe A, Skrede OJ, De Raedt S, et al: A clinical decision support system optimising adjuvant chemotherapy for colorectal cancers by integrating deep learning and pathological staging markers: A development and validation study. Lancet Oncol 23:1221-1232, 2022.
EXPERT POINT OF VIEW
Invited study discussant, Arsela Prelaj, MD, PhD, of the Fondazione IRCCS Istituto Nazionale dei Tumori, Milan, Italy, congratulated the authors of the interim report of a Dutch nationwide validation of the Combined Analysis of Pathology and Artificial Intelligence (CAPAI) model in patients with locally advanced colon cancer treated without adjuvant chemotherapy, calling it “fantastic” work in the digital oncology space. She emphasized its methodologic and scientific value. “The CAPAI model represents a breakthrough in artificial intelligence/pathologist collaboration, fusing AI morphology with classic pathologic expertise,” she commented. “This combination improves interpretability and aligns with the emerging ‘augmented pathologist’ model.”
Arsela Prelaj, MD, PhD
Dr. Prelaj highlighted that previous studies showed that “AI combined with pathology outperformed guideline staging and identified around 40% of patients [with CAPAI-defined low-risk disease who might have] safely avoided chemotherapy.”1 She praised the current analysis for demonstrating the model’s prognostic power independent of chemotherapy, confirming that approximately 47% of untreated patients could indeed be spared chemotherapy.
Key limitations and areas for further investigation, according to Dr. Prelaj, include selection bias from potentially having withheld chemotherapy in perceived low-risk patients, the retrospective vs prospective study design, a lack of molecular integration (eg, circulating tumor DNA, RAS/BRAF, and immune-infiltration markers), limited generalizability beyond Scandinavian and Dutch cohorts, and clinical implementation challenges (eg, workflow integration, standardization, and regulatory readiness). “But this is a real opportunity to put biomarkers like this in decision-making,” she added.
Dr. Prelaj shared these closing remarks: “In an ESMO EBAI [ESMO Scale of Minimum Requirements for AI Biomarkers in Oncology] paper that will be published in Annals of Oncology…, these types of biomarkers [C2: those informing treatment decisions] need prospective validation. Consider novel foundation models as technology to see [whether they can improve these results]. [Until prospective validation], another interesting suggestion can be…[to] use CAPAI to prescreen patients who can be monitored through circulating tumor DNA; this can be another way to rationalize cost.”
DISCLOSURE: Dr. Prelaj has served as an invited speaker for AstraZeneca, Daiichi Sankyo, Gilead Sciences, IQVIA, Janssen, Lilly, MEDSIR, Novartis, Pfizer, and Roche; has served on advisory boards for Amgen, AstraZeneca, Bayer, BMS, Johnson & Johnson, MSD, and Pfizer; has served as a local principal investigator for AstraZeneca, Spectrum, Bayer, BMS, Lilly, MSD, and Roche; and has nonfinancial interests as Project Lead for APOLLO 11 and I3LUNG as well as President of the European Interdisciplinary Society of AI in Cancer Research.
REFERENCE
1. Kleppe A, Skrede OJ, De Raedt S, et al: A clinical decision support system optimising adjuvant chemotherapy for colorectal cancers by integrating deep learning and pathological staging markers: A development and validation study. Lancet Oncol 23:1221-1232, 2022.
