AI-Derived Electronic Tumor Marker For Cancer Antigen 19-9 Nonproducers With Pancreatic Ductal Adenocarcinoma.
[IMPORTANCE] Cancer antigen 19-9 (CA19-9) is used to assess treatment response among patients with pancreatic ductal adenocarcinoma (PDAC); however, nearly 30% of patients with PDAC do not produce ele
- p-value P = .006
- p-value P < .001
- 95% CI 1.60-15.66
- 연구 설계 cohort study
APA
Thalji SZ, Aldakkak M, et al. (2026). AI-Derived Electronic Tumor Marker For Cancer Antigen 19-9 Nonproducers With Pancreatic Ductal Adenocarcinoma.. JAMA surgery. https://doi.org/10.1001/jamasurg.2026.0291
MLA
Thalji SZ, et al.. "AI-Derived Electronic Tumor Marker For Cancer Antigen 19-9 Nonproducers With Pancreatic Ductal Adenocarcinoma.." JAMA surgery, 2026.
PMID
41848749
Abstract
[IMPORTANCE] Cancer antigen 19-9 (CA19-9) is used to assess treatment response among patients with pancreatic ductal adenocarcinoma (PDAC); however, nearly 30% of patients with PDAC do not produce elevated CA19-9.
[OBJECTIVE] To develop, validate, and apply an electronic tumor marker (e19-9) derived from routine laboratory data available in the electronic health record to assess treatment response and predict outcomes among patients with PDAC who do not produce CA19-9.
[DESIGN, SETTING, AND PARTICIPANTS] In this cohort study, an artificial intelligence (AI) model was trained using routinely collected serum laboratory data from patients with PDAC and elevated CA19-9. The model was externally validated and then applied to a separate cohort of CA19-9 nonproducers. Model development and internal testing were conducted at a single institution using patient data from 2010 to 2022. External validation used a deidentified patient network across 58 health care organizations over the same period. The training cohort included 3239 patients with pancreatic cancer and elevated CA19-9. The external validation cohort included 4384 similar patients. The model was applied to 121 patients with resectable or borderline resectable PDAC who did not produce elevated CA19-9 and received neoadjuvant therapy with curative intent. These data were analyzed from November 2021 through March 2025.
[MAIN OUTCOMES AND MEASURES] Model performance was assessed using root mean square error and R2. Clinical outcomes included completion of all neoadjuvant treatment and surgery, metastatic progression, and overall survival (OS).
[RESULTS] The final fitted model demonstrated stable performance across both internal and external validation cohorts. Among 121 patients (59 female and 62 male) with localized PDAC who did not produce elevated CA19-9, a 50% or more decline in e19-9 (area under the curve [AUC], 0.79) and e19-9 level of less than 100 (AUC, 0.84) were objectively determined cut points associated with prognosis. A total of 93 patients (77%) completed all planned neoadjuvant therapy and surgery. A 50% or more decline in e19-9 levels and an e19-9 level less than 100 was associated with completion of all intended therapy (odds ratio [OR], 5.00; 95% CI, 1.60-15.66; P = .006 and OR, 19.31; 95% CI, 5.80-64.26; P < .001). An e19-9 level less than 100 was independently associated with OS (hazard ratio, 0.49; 95% CI, 0.25-0.97; P = .04).
[CONCLUSIONS AND RELEVANCE] In this study, e19-9 was a noninvasive AI-derived marker that may provide accurate and relevant information to assess treatment response for the approximately 30% of patients with PDAC who do not produce CA19-9 at elevated levels. The development and validation of scalable, noninvasive screening methods using machine-learning algorithms may pave the way for early detection, prognostication, and treatment of cancers.
[OBJECTIVE] To develop, validate, and apply an electronic tumor marker (e19-9) derived from routine laboratory data available in the electronic health record to assess treatment response and predict outcomes among patients with PDAC who do not produce CA19-9.
[DESIGN, SETTING, AND PARTICIPANTS] In this cohort study, an artificial intelligence (AI) model was trained using routinely collected serum laboratory data from patients with PDAC and elevated CA19-9. The model was externally validated and then applied to a separate cohort of CA19-9 nonproducers. Model development and internal testing were conducted at a single institution using patient data from 2010 to 2022. External validation used a deidentified patient network across 58 health care organizations over the same period. The training cohort included 3239 patients with pancreatic cancer and elevated CA19-9. The external validation cohort included 4384 similar patients. The model was applied to 121 patients with resectable or borderline resectable PDAC who did not produce elevated CA19-9 and received neoadjuvant therapy with curative intent. These data were analyzed from November 2021 through March 2025.
[MAIN OUTCOMES AND MEASURES] Model performance was assessed using root mean square error and R2. Clinical outcomes included completion of all neoadjuvant treatment and surgery, metastatic progression, and overall survival (OS).
[RESULTS] The final fitted model demonstrated stable performance across both internal and external validation cohorts. Among 121 patients (59 female and 62 male) with localized PDAC who did not produce elevated CA19-9, a 50% or more decline in e19-9 (area under the curve [AUC], 0.79) and e19-9 level of less than 100 (AUC, 0.84) were objectively determined cut points associated with prognosis. A total of 93 patients (77%) completed all planned neoadjuvant therapy and surgery. A 50% or more decline in e19-9 levels and an e19-9 level less than 100 was associated with completion of all intended therapy (odds ratio [OR], 5.00; 95% CI, 1.60-15.66; P = .006 and OR, 19.31; 95% CI, 5.80-64.26; P < .001). An e19-9 level less than 100 was independently associated with OS (hazard ratio, 0.49; 95% CI, 0.25-0.97; P = .04).
[CONCLUSIONS AND RELEVANCE] In this study, e19-9 was a noninvasive AI-derived marker that may provide accurate and relevant information to assess treatment response for the approximately 30% of patients with PDAC who do not produce CA19-9 at elevated levels. The development and validation of scalable, noninvasive screening methods using machine-learning algorithms may pave the way for early detection, prognostication, and treatment of cancers.