2
Díaz Cantón E. Exacꢀtud, estadiꢁcación y decisión: la inteligencia arꢀꢁcial frente al cáncer esofagogástrico. Rev Argent Cir. 2026;118(2):e-ededc02
guías. Medida así, con rigor y contra el patrón correcto, herramientas de inteligencia arꢀꢁcial generaꢀva basadas
en modelos de lenguaje de gran escala como apoyo en
la exacꢀtud de la IA deja de ser una promesa para con-
la búsqueda y el cotejo de la bibliograꢃa, la organización
verꢀrse en una herramienta; medida contra su propia
del texto y la edición de esꢀlo. Todas las aꢁrmaciones, las
elocuencia, sigue siendo un espejismo.
referencias y los datos cuanꢀtaꢀvos fueron veriꢁcados por
el autor contra las fuentes primarias citadas. El contenido
conceptual, las interpretaciones y las conclusiones son de su
exclusiva responsabilidad; la inteligencia arꢀꢁcial no ꢁgura
como autora ni asume responsabilidad alguna sobre el
contenido.
Declaración sobre el uso de inteligencia arꢀꢁcial: El autor
declara que en la preparación de este editorial uꢀlizó
■
ENGLISH VERSION
In gastroesophageal cancer, staging is not me- From staging to decision-making
rely a descripꢀve exercise: it guides high-impact, oꢁen
conꢂicꢀng decisions. The paper “Accuracy of arꢀficial
When the focus shiꢁs from staging to recom-
intelligence for staging and decision-making in esopha- mending, concordance becomes more variable. For
1
geal and gastric cancer” raises the perꢀnent quesꢀon predicꢀng tumor board procedural recommendaꢀons,
of how accurate arꢀficial intelligence (AI) is. More im- treatment line congruence of up to 86% has been re-
4
portantly, it requires specifying the standard against ported with some models , but the same ecosystem
which this accuracy is measured.
generates one-third of non-concordant recommenda-
The most direct evidence is encouraging, ꢀons when queried about open oncology treatment,
5
though condiꢀonal. Yao et al. compared the perfor- with 12.5% of responses being hallucinaꢀons . Table
mance of three locally deployed LLMs with that of cli- 1 summarizes this heterogeneous evidence, which
nicians in preoperaꢀve esophageal cancer staging using should be read with cauꢀon.
free-text radiology reports. The gold standard was de-
The objecꢀve is not to discredit the tool, but
rived from postoperaꢀve pathological staging. When rather to contextualize its use. Arꢀficial intelligence can
the authors tested three prompꢀng strategies—zero- improve staging and provide explicit, verifiable reaso-
shot, chain-of-thought, and an interpretable reasoning ning, especially when guided by structured prompts.
method—they observed that the structured prompt However, the ethical, legal, and human responsibility
not only improved accuracy but also provided an au- for the decision cannot be delegated; it remains with
2
ditable reasoning trace . Therefore, the accuracy of AI the surgeon and the oncologist, who must verify each
can be conꢀngent on the method of querying, and it output against the individual paꢀent, their pathology,
should be evaluated against the pathological gold stan- and current guidelines. When evaluated with such rigor
dard rather than the plausibility of a well-wriꢃen text.
■
TABLE 1
Accuracy and concordance of AI in cancer staging
Why does gastroesophageal cancer demand parꢀcular
cauꢀon?
and decision-making.
Domain
Model(s)
Primary outcome
Source
[2]
Few scenarios impose such a heavy penalty
for a staging error. The management of the disease de-
pends on subtle disꢀncꢀons—tumor locaꢀon relaꢀve to
the gastroesophageal juncꢀon, Siewert classificaꢀon,
and histological differences between adenocarcinoma
and squamous cell carcinoma—that guide different
perioperaꢀve or chemoradiotherapy regimens. In addi-
TNM staging of
High accuracy;
structured reasoning
(CoT / interpretable
reasoning)
esophageal cancer
INF-72B,
based on radiology Qwen2.5-72B,
reports (gold
LLaMA3.1-70B
standard: pathology)
outperformed zero-shot
Treatment line
congruence of up
to 86%; jusꢀfiable
recommendaꢀons up
to 98%.
Predicꢀon of tumor
board procedural
recommendaꢀons
(head and neck)
Various LLMs
[4]
[5]
ꢀ
on, certain biomarkers, such as HER2, PD-L1, micro-
satellite instability, and CLDN18.2, condiꢀon systemic
Open oncology
recommendaꢀons
vs. NCCN guidelines
3
therapy . A model confounding a cardia tumor with that
3
4.3% with ≥ 1
of the gastroesophageal juncꢀon—or one that ignores
a biomarker—spreads the error throughout the enꢀre
treatment plan. The “accuracy” claimed by this study
must therefore be assessed within the framework of
current guidelines and in relaꢀon to the pathological
diagnosis.
GPT-3.5
nonconcordant opꢀon;
12.5% “hallucinaꢀons”
(
breast cancer,
prostate cancer, lung
cancer)
Note: These values are derived from studies using different methodolo-
gies, languages, and populaꢀons and do not consꢀtute a direct compari-
son between models. CoT: chain-of-thought; NCCN: Naꢀonal Comprehen-
sive Cancer Network; TNM: tumor-node-metastasis