Banca de DEFESA: RAYONARA MEDEIROS DE AZEVEDO

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
STUDENT : RAYONARA MEDEIROS DE AZEVEDO
DATE: 01/06/2026
TIME: 14:00
LOCAL: Departamento de Enfermagem
TITLE:

Prediction of Cervical Cytopathological Changes Using Artificial Intelligence in Primary Health Care

KEY WORDS:

Nursing; Cervical neoplasms; Machine learning; Artificial intelligence; Screening; Primary health care.

PAGES: 137
BIG AREA: Ciências da Saúde
AREA: Enfermagem
SUMMARY:

Cervical cancer remains a significant public health challenge in Brazil,

primarily due to the low effectiveness of opportunistic cytopathological screening and

regional disparities in access. The transition to molecular HPV DNA testing as

a primary strategy, combined with the incorporation of Artificial Intelligence technologies,

represents a promising strategy for improving early detection and risk stratification.

Thus, this study aims to develop a machine learning model

for predicting cervical cytopathological abnormalities based on

sociodemographic, clinical, and reproductive data. This is a cross-sectional, diagnostic prediction

study with a quantitative approach, conducted in primary care units

in four municipalities in the state of Rio Grande do Norte, with data collection carried out between June and

December 2025. The sample comprised 483 women aged 25 to 64 years who underwent a

structured interview and standardized cytopathological examination. Descriptive,

bivariate (chi-square/Fisher), and multivariate logistic regression analyses were performed. For predictive modeling,

the data were partitioned into training (70%) and testing (30%) sets with stratification. Feature selection

employed the Boruta algorithm, and five classifiers were trained and compared:

Random Forest, XGBoost, LightGBM, CatBoost, and TabPFN, with Bayesian optimization of

hyperparameters via Optuna and stratified cross-validation (k=5). The interpretability of the

selected model was assessed using the SHAP method. The study was approved by the

Research Ethics Committee under opinion no. 7.296.33. Of the 483 samples evaluated, 99.17% were

satisfactory, with an overall prevalence of cellular atypia in 14.91% (n=72) of the samples and

a predominance of ASC-US in 9.52%. The following were identified as independent predictors of

cytopathological abnormalities: treatment for vaginal infection in the past six months (adjusted OR=2.39;

95% CI 1.31–4.33; p=0.005), parity (Adj. OR=3.20; p=0.032), absence of prior treatment

for HPV (Adj. OR=2.67; p=0.047), current smoking (Adj. OR=2.67; p=0.017), and age up to 42 years

(Adj. PR=1.92; p=0.019). Boruta confirmed five predictors (smoking, recent treatment for

vaginal infection, income up to one minimum wage, parity, and vaginal delivery). The five models

showed accuracy around 0.841, with TabPFN reaching 0.848. CatBoost achieved

the highest discriminative power (AUC-ROC=0.634), followed by Random Forest (0.618) and

XGBoost (0.617). The SHAP analysis of CatBoost highlighted a greater predictive contribution from

treatment for recent vaginal infection (≈0.180), median income (≈0.140), and current smoking

(≈0.105), partially corroborating Boruta’s selection. The developed model demonstrated

moderate discriminative performance. The convergence among the predictors identified by

logistic regression, Boruta, and SHAP reinforces the epidemiological consistency of the findings

and signals the potential of machine learning models as a complementary tool

for risk stratification in reflex cytology, requiring sample expansion and the incorporation

of additional variables for future refinement.

COMMITTEE MEMBERS:
Interna - 3149773 - DANDARA NAYARA AZEVEDO DANTAS
Externa à Instituição - DANYELLE LEONETTE ARAUJO DOS SANTOS
Externo ao Programa - 3050428 - DIEGO BONFADA - nullExterna ao Programa - 3995493 - MARCELLY SANTOS COSSI - nullPresidente - 1054749 - MARIA ISABEL DA CONCEICAO DIAS FERNANDES
Externo à Instituição - MARQUIONY MARQUES DOS SANTOS - UERN