Artigos

INTRODUCTION
Health systems grounded in Primary Health Care (PHC) principles demonstrate enhanced equity, efficiency, and sustainability, which improved population health outcomes 1-3. In Brazil, despite the nationwide expansion of PHC services 1, persistent challenges include uneven service quality, fragmented care, and limited integration of health technologies, hindering the full realization of PHC goals 4, 5. Rio de Janeiro exemplifies these dynamics: over 15 years of reform have prioritized the Family Health Strategy (ESF), increasing coverage from 3.5% (2008) to 70% (2025) alongside the establishment of Family Health Clinics (CFs), expanded service portfolios, and residency programs to strengthen care delivery 6-9. However, systematic evaluations of user experiences post-reform remain sparse, particularly regarding how PHC attributes align with population needs.
The Primary Care Assessment Tool (PCATool) is a globally validated instrument for evaluating PHC performance, enabling cross-context comparisons through its adaptable framework 10-13. In Rio de Janeiro, a 2014 PCATool-Brazil study highlighted disparities in user perceptions: children rated PHC services more favorably than adults, with persistent weaknesses in accessibility but strengths in longitudinality and care coordination 14. While such studies provide a deep understanding of PHC quality, the PCATool’s length and complexity often hinder its practical use in day-to-day service monitoring, particularly in resource-constrained environments. Furthermore, no research has used machine learning (ML) to develop predictive models of PCATool outcomes or conducted external validation using longitudinal data, limiting the ability to forecast PHC performance and inform targeted interventions.
Machine learning offers transformative potential for healthcare analytics, identifying complex, nonlinear patterns in multidimensional datasets 15—capabilities underutilized in PHC evaluation. If ML models can achieve robust predictive performance using a subset of key variables (e.g., user-reported health status and service utilization patterns), they could streamline PHC assessments, bypassing the need for exhaustive PCATool questionnaires. This approach could enable simpler, faster implementation in routine care settings, facilitating early identification of service gaps and timely interventions. Yet, no studies have applied ML to predict PCATool scores or validated such models with external, decade-spanning datasets, a critical gap given the dynamic nature of health systems undergoing rapid scaling.
To address these gaps, this study aims to evaluate the predictive performance of machine learning algorithms in estimating PCATool scores, leveraging two datasets spanning 10 years. By integrating ML into PHC assessment, this research advances methodological innovation in health systems evaluation and explores its potential to enhance monitoring efficiency. If successful, these models could reduce dependence on lengthy PCATool surveys by prioritizing key variables associated with user health and service engagement, enabling more rapid and scalable assessments. The inclusion of external validation using decade-long data strengthens the study’s robustness, addressing a critical limitation of prior predictive studies in low-resource settings.

METHODS:
This study adheres to the TRIPOD+ AI guidelines for reporting clinical prediction models that incorporate regression and machine learning methodologies 16.

Data:
Data were sourced from two cross-sectional studies performed in Rio de Janeiro in 2014 14 and 2024 with similar study designs. The studies involved independent random samples of service users from ten planning areas in the city. The data was collected through structured interviews conducted by trained interviewers. The studies applied the PCATool-Brazil to assess the extent to which healthcare services adhered to Primary Healthcare principles, considering sociodemographic characteristics and self-reported morbidity. Ethical approval was obtained from the City of Rio de Janeiro Department of Health (SMS-RJ) Ethics Committee under number 133/13 for the 2014 study and Federal University of Porto Alegre Ethics Committee number 77802624.0.0000.5347/2024 for the 2014 study. Further details on sampling and methodological aspects are available in official publications 9, 14. Data collection occurred between January and June 2014 and between March to April 2024.

Participants:
A total of 3,530 adults and 3,145 children participated in the study in 2014, while 968 adults and 985 children participated in 2024. Eligibility criteria included service users aged 18 or older for the adult sample and ?12 years for children, who had attended at least two medical visits at the same healthcare unit in the previous two years. Participants were recruited upon exiting a medical appointment at the healthcare unit on the interview day. Healthcare units included in the study were required to have implemented the Family Health strategy for a minimum of six months, according to the Secretaria Municipal de Saúde do Rio de Janeiro (SMS-RJ) registry from July 2013. Individuals unable to complete the questionnaire due to cognitive or physical impairments, as well as those who had not met the minimum consultation requirement, were excluded. For the child, the caregiver answered the questionnaire at the time of the consultation.

Sample Size:
For 2014, the sample size was determined to facilitate comparisons of Primary Healthcare orientation across Family Clinics (CF), Healthcare Centers (HCC-A), and Mixed Healthcare Centers (HCC-B), based on assessments from both adult and child users. The calculation assumed a minimum difference of 0.5 in the average overall score among the three healthcare service types. A significance level of 5% was applied, with statistical power set at 90% for adults and 80% for children. The estimated required sample included 2,884 adult and 2,600 users. To compensate for potential losses, an additional 10% was included.
For 2024, a 95% confidence level and a 4% margin of error (d) were considered for the estimates, taking into account that, in addition to the questionnaire scores generated by the PCATool, various frequency-based questions would also be analyzed. Due to the lack of recent studies, the most conservative scenario was assumed, with p=q=0.5. Based on this, the estimated sample size was 1,034, including a 10% allowance for potential losses for both adult and children sample.

Outcomes:
The study's outcome variable was the: i) General and ii) Essential scores of the PCATool Brazil for adults and children. These scores, along with the Primary Healthcare (PHC) attribute scores, were calculated following the instrument's official manual 17 and presented on a scale from 0 to 10. The item scores were derived by converting the original scale (ranging from 1 to 4) to align with the 0-to-10 scale used for the final scores. A score above 6.6 indicates a high quality of healthcare for the respective item or attribute. Consequently, scores were categorized as "High Score" (above 6.6) and "Low Score" (6.6 or below).

Predictors
Forty-five predictors for adults and forty predictors for children were selected based on prior literature and expert knowledge. The predictors were present in both cross-sectional studies. A detailed list of these variables is provided in Supplemental Material S1. Variables include geographical region (Rio de Janeiro, other regions), sex (female or male), race/ethnicity (White, Black, Brown, Yellow, Indigenous), and age (numeric). Additional variables include family structure, socioeconomic variables, health conditions, self-reported health, health service access, and last medical visit.

Data Preparation
Data preprocessing was conducted using R version 4.3.2 software (R Core Team, Vienna, Austria). This involved data cleaning, handling missing values, integrating datasets, and transforming variables. Categorical variables with more than two categories were converted into binary format using one-hot encoding, while numerical variables were standardized as Z-scores.

Missing Data
Missing values in the predictors were imputed using the Multivariate Imputation by Chained Equations (MICE) algorithm with a random forest approach.

Statistical Methods:
Descriptive analyses were performed using Stata 16.0. Subsequent analyses were conducted in Python version 3.11.7. The 2014 dataset was split (80:20) into training and test sets, while the 2024 dataset was used for external validation. Five machine learning (ML) algorithms were evaluated: Gradient Boosting Classifier, AdaBoost Classifier, Extreme Gradient Boosting, Light Gradient Boosting Machine, and CatBoost Classifier. Class imbalance adjustments were not applied. Hyperparameter tuning was performed using 10-fold cross-validation with 50 iterations, optimizing the Receiver Operating Characteristic (ROC) curve (AUC). Model performance was compared using AUC values, with 95% confidence intervals generated via 10,000 bootstrap resamples. Additional performance metrics, including accuracy, recall, precision, and F1-score, were computed. Differences between ROC curves were analyzed using the DeLong test in R software.
To interpret the most effective predictive model, Shapley values were calculated using the SHapley Additive exPlanations (SHAP) algorithm 18, identifying the contribution of each variable to the model’s predictions. Model fairness was assessed through stratified analyses by sex and race using the algorithm with the highest AUC. The significance threshold for all statistical analyses was set at 0.05.


?
RESULTS:
Of 3,530 adults and 3145 children who participated in 2014, 3,528 (99.9%) and 3,116 (99.1%) have respectively full data from PCATool and were included in the study. Of 968 adults and 985 children who participated in the 2024 study, 889 (91.8%) and 956 (97.1%) have respectively full data from PCATool and were included in the study. The adult sample was constituted mainly of females (79.3%) and Brown skin color (44.3%) individuals, mean age of 46.9 (SD=17.2), and from SES C2 (37.1%). The child sample was constituted mainly of females (50.1%) and Brown skin color (46.5%) individuals, from SES B1 (42.3%). Regarding adults, the prevalence of a High Score was 34.9% for PCATool’s general score and 39.5% for the essential score. Regarding children, the prevalence of a High Score was 47.2% for PCATool’s general score and 54.3% for the essential score.
The performance of the machine learning models for predicting PCATool general and essential scores in adults and children is presented in Supplemental Tables 2 to 5 and Supplemental Figures 1 and 2. A summarization of the best models is displayed in Table 1. For the PCATool’s general scores in adults, the Random Forest Classifier demonstrated the best overall performance, achieving a moderate performance and the most balanced trade-off across the evaluation parameters. In the test set was the Random Forest Classifier, achieving an AUC of 0.68 (95% CI [0.64–0.72]), with a recall of 0.66 and a precision of 0.41. In the validation set, it maintained the performance (AUC = 0.66, 95% CI [0.62–0.69]), with balanced recall (0.59) and precision (0.62), making it the most reliable model for this outcome. For the essential scores in adults, the best model was also the Random Forest Classifier, with moderate performance and considering the balanced trade-off across all evaluation parameters. The model outperformed others in the test set (AUC = 0.68, 95% CI [0.63–0.72], recall = 0.64, precision = 0.45). Its performance declined slightly in the validation set (AUC = 0.65, 95% CI [0.62–0.69]), with a recall of 0.58 and a precision of 0.68).
For the PCATool’s general scores in children, the Extreme Gradient Boost demonstrated the best overall performance, achieving a moderate performance and the most balanced trade-off across the evaluation parameters. In the test set it achieved an AUC of 0.65 (95% CI [0.60–0.69]), with a recall of 0.94 and a precision of 0.43. In the validation set, it sight decreased the performance (AUC = 0.63, 95% CI [0.59–0.66]), with good recall (0.92) and precision (0.73). For the essential scores in children, the best model was the Extra Threes Classifier, with moderate performance and considering the balanced trade-off across all evaluation parameters. The model presented in the test set an AUC = 0.69, (95% CI [0.65–0.73], recall = 0.60, precision = 0.62). Its performance declined slightly in the validation set (AUC = 0.66, 95% CI [0.62–0.70]), with a recall of 0.50 and a precision of 0.88).
Supplemental Table 6 presents the fairness estimations of the best models on the test set. Race-based analysis revealed that Brown individuals had higher performances for PCATool general scores in adults and Black individuals had higher performances for PCATool essential scores in adults. Gender-based analysis revealed similar AUC and F1-score for males and females, though males had a higher recall and lower precision in the adults set. In the children set, the race-based analysis revealed that black individuals had higher performance in general and essential scores. Gender-based analysis for children set revealed that males performed better in the general score while females performed better in the essential score.
Regarding interpreting algorithmic decision-making, we developed the Shapley values plot for the best models for each score (general or essential) and age (adult and children) (Figure 1). The analysis reveals that the key variable influencing the model's classification was 'Consulting with the same doctor’, which was the first variable for the three models. Other significant factors included having a high number of consultations in recent years and using the service for more than two years.

DISCUSSION
This study demonstrates the potential of machine learning to enhance Primary Healthcare (PHC) evaluation by predicting PCATool scores with moderate performance. The Random Forest model emerged as the most reliable classifier for adults, maintaining stable performance across both test and validation datasets. On the other hand, Extreme Gradient Boosting and Extra Tree Classifier were the best models for predicting the general and essential scores for children, respectively. Notably, key predictors such as continuity of care with the same doctor, and frequency of service utilization, were consistently important to predicting the PCATool scores. These results highlight longitudinality as a key factor reinforcing the essential role of sustained patient-provider relationships in PHC 19, 20. The consistent significance of longitudinal care aligns with previous evidence emphasizing its impact on service accessibility, care coordination, and patient-centered outcomes 21. This finding emphasizes the need for healthcare policies that prioritize long-term patient engagement and continuity of care to improve PHC quality.
Our findings suggest that machine learning could be explored as a complementary tool for PHC assessment, potentially reducing reliance on lengthy surveys while still capturing relevant aspects of service quality and patient experience 22. However, the model’s moderate performance indicates that further refinements are necessary to enhance predictive accuracy and reliability. Additionally, fairness analyses revealed a slight performance disparity across racial groups and gender groups, which was observed in other studies 23, 24 needing caution 25, 26. This underscores the need for ongoing evaluation and model refinement to ensure equitable application across diverse populations and to mitigate potential biases 24. In this sense, further strategies could be explored to enhance predictive accuracy, such as incorporating new explanatory variables – when available – related to service organization, provider workload, or geographic accessibility. Likewise, evaluating more complex model architectures, including deep neural networks or hybrid ensemble – deep learning approaches, may uncover nonlinear patterns. These enhancements could strengthen model robustness and improve performance across temporally and contextually diverse PHC settings. Moreover, the fairness findings described – such as higher performance among black individuals for the essential score in adults – deserve deeper consideration, as they may reflect underlying data biases, differential healthcare utilization patterns not fully captured by the predictors, or broader sociodemographic structures shaping interactions with PHC services. Examining these mechanisms in future work will be essential to improving fairness and equity in predictive models.
An important factor influencing our outcome in the validation test may be the temporal gap between training and validation datasets. A meta-analysis investigating chronic obstructive pulmonary disease observed that most studies that performed external validation had a significant decrease in performance 27. Given that in our study the training and test sets were derived from 2014 data, the models likely learned associations that were relevant at the time but less applicable to the 2024 dataset. Changes in healthcare policies 1, 28, service accessibility 29, and patient behaviors over a decade could have altered the underlying distribution of key predictors, diminishing model performance. Despite these challenges, the findings reinforce the feasibility of machine learning in PHC evaluation and highlight avenues for model refinement. Incorporating longitudinal learning approaches, domain adaptation techniques, and fairness-aware modeling strategies could enhance predictive robustness and mitigate biases.
The integration of machine learning into Primary Healthcare (PHC) has shown promising potential in optimizing healthcare delivery, improving resource allocation, and enhancing patient outcomes 30. Beyond its application in diagnosis 31, 32, and its evaluation of service quality through tools like PCATool, machine learning has also been effectively employed in operational management, such as predicting patient no-shows 30. A recent study demonstrated that machine learning-based optimization models could significantly mitigate the impact of missed appointments by analyzing key factors such as patient demographics, appointment details, and external conditions 30. However, the literature is still scarce when it comes to applying machine learning approaches specifically to public health settings, particularly in the evaluation and enhancement of PHC services. Most studies focus on hospital-based care 33, and specialized medicine 34, 35, leaving a gap in research exploring how predictive models can support decision-making, improve care continuity, and identify systemic inefficiencies in PHC.
We are aware that our study has limitations that need to be discussed. Machine learning models were trained on two specific datasets from Rio de Janeiro, which may limit their generalizability to other populations, healthcare settings, or geographic regions 36. To address this challenge, transfer learning strategies have been successfully employed and hold promise for future applications 36. Although fairness analyses were conducted, residual biases in model predictions were observed, particularly concerning racial and sex groups. This highlights the need for continuous monitoring, bias mitigation strategies, and fairness-aware modeling approaches to enhance equity in predictions 37. While the models demonstrated the potential to streamline PHC assessment, they should not be viewed as replacements for traditional survey-based methods. The reduction in survey length, while beneficial for efficiency, may lead to information loss, particularly in capturing nuanced aspects of patient experience and service quality. Further research should investigate the trade-offs between survey reduction and the retention of critical insights. Additionally, implementing predictive tools in real-world PHC settings requires careful planning, transparent validation, and sensitivity to ethical concerns related to data use, explainability, and clinician trust.
A further consideration involves the steps required to translate these findings into routine PHC practice. Future implementation will require external validation in other cities and regions to ensure generalizability across diverse organizational structures and population profiles. Developing user-friendly interfaces that allow managers and frontline professionals to interpret predictions is also essential for practical integration. Ethical challenges – particularly those related to transparency, accountability, data privacy, and potential resistance or skepticism among healthcare professionals – must be anticipated to promote trustworthy use of machine learning tools. Addressing these issues is a critical step toward bridging the gap between methodological innovation and sustainable adoption of predictive systems in PHC.
While model performance metrics indicated satisfactory predictive capability, real-world implementation may introduce additional challenges, including ethical concerns, interpretability issues, and resistance from healthcare providers. Ensuring that these models are effectively integrated into decision-making processes will require interdisciplinary collaboration, user-centered design, and transparent validation efforts. In addition to these challenges, practical constraints such as limited digital infrastructure, variable data quality, restrictions on workforce training, and potential workflow disruptions must be considered when integrating predictive tools into PHC routines. Implementing machine learning based tools in routine care requires not only technical robustness but also strategies to support adoption, such as clear governance procedures, continuous model monitoring, and mechanisms to ensure accountability and trust among providers and users. Understanding and addressing these real-world barriers remain essential to translating predictive performance into meaningful improvements in PHC delivery.
Author contributions:
Luiz Alexandre Chisini: conceptualization, methodology, formal analysis, data curation, writing – original draft, visualization. Otávio Pereira D'Avila: conceptualization, methodology, writing – review & editing, supervision. Mauro Cardoso Ribeiro: methodology, writing – review. Luiz Felipe Pinto: conceptualization, resources, writing – review & editing. Erno Harzheim: conceptualization, supervision, resources, writing – review & editing.

DATA AVAILABILITY STATEMENT

Os dados de pesquisa estão disponíveis mediante solicitação ao autor de correspondência.

REFERENCES

1. D'Avila OP, Chisini LA, Costa FDS, Cademartori MG, Cleff LB, Castilhos ED. Use of Health Services and Family Health Strategy Households Population Coverage in Brazil. Cien Saude Colet. 2021;26(9):3955-64.doi:10.1590/1413-81232021269.11782021
2. Cassady CE, Starfield B, Hurtado MP, Berk RA, Nanda JP, Friedenberg LA. Measuring consumer experiences with primary care. Pediatrics. 2000;105(4 Pt 2):998-1003
3. Macinko J, Starfield B, Shi L. The contribution of primary care systems to health outcomes within Organization for Economic Cooperation and Development (OECD) countries, 1970-1998. Health Serv Res. 2003;38(3):831-65.doi:10.1111/1475-6773.00149
4. Macinko J, Guanais FC, de Fatima M, de Souza M. Evaluation of the impact of the Family Health Program on infant mortality in Brazil, 1990-2002. J Epidemiol Community Health. 2006;60(1):13-9.doi:10.1136/jech.2005.038323
5. Hone T, Been JV, Saraceni V, Coeli CM, Trajman A, Rasella D, et al. Associations between primary healthcare and infant health outcomes: a cohort analysis of low-income mothers in Rio de Janeiro, Brazil. Lancet Reg Health Am. 2023;22:100519.doi:10.1016/j.lana.2023.100519
6. Soranz D, Pinto LF, Penna GO. Themes and Reform of Primary Health Care (RCAPS) in the city of Rio de Janeiro, Brazil. Cien Saude Colet. 2016;21(5):1327-38.doi:10.1590/1413-81232015215.01022016
7. Justino AL, Oliver LL, Melo TP. Implementation of the Residency Program in Family and Community Medicine of the Rio de Janeiro Municipal Health Department, Brazil. Cien Saude Colet. 2016;21(5):1471-80.doi:10.1590/1413-81232015215.04342016
8. Pinto LF, Caldas ALF. Fifteen years of the Primary Health Care Reform (RCAPS) in Rio de Janeiro, Brazil. . Cien Saude Colet. 2025;30(7):1-16 e04112025
9. D’Avila OP, Chisini LA, Ribeiro MC, Meira-Silva VST, Mathuiy YR, Moura LJN, et al. Evaluation of Primary Health Care in Rio de Janeiro: the experience of patients fifteen years after the Reform. Cien Saude Colet. 2025;evaluable at: http://cienciaesaudecoletiva.com.br/artigos/avaliacao-da-atencao-primaria-a-saude-no-rio-de-janeiro-experiencia-de-usuarios-apos-quinze-anos-da-reforma/19544
10. Shi L, Starfield B, Xu J. Validating the adult primary care assessment tool. J Fam Pract 2001;50(2):161-75
11. Menegazzo GR, Fagundes MLB, Amaral Junior OLD, Bastos LF, Cunha ARD, Neves M, et al. Psychometric properties of the reduced version of the Primary Care Assessment Tool (PCATool). Rev Bras Epidemiol. 2024;27:e240057.doi:10.1590/1980-549720240057
12. Carvalho FC, Bernal RTI, Perillo RD, Malta DC. Association between positive assessment of Primary Health Care, sociodemographic characteristics and comorbidities in Brazil. Rev Bras Epidemiol. 2022;25:e220023.doi:10.1590/1980-549720220023.2
13. Guimaraes MA, Fattori A, Coimbra AMV. "PCATool version to professionals in the primary care of the elderly": adaptation, content analysis and first results. Cien Saude Colet. 2022;27(7):2911-9.doi:10.1590/1413-81232022277.19292021
14. Harzheim E, Pinto LF, Hauser L, Soranz D. Assessment of child and adult users of the degree of orientation of Primary Healthcare in the city of Rio de Janeiro, Brazil. Cien Saude Colet. 2016;21(5):1399-408.doi:10.1590/1413-81232015215.26672015
15. Rajkomar A, Dean J, Kohane I. Machine Learning in Medicine. N Engl J Med. 2019;380(14):1347-58.doi:10.1056/NEJMra1814259
16. Collins GS, Moons KGM, Dhiman P, Riley RD, Beam AL, Van Calster B, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:e078378.doi:10.1136/bmj-2023-078378
17. Brasil. Ministério da saúde. secretaria de atenção Primária à saúde. Departamento de saúde da Família.
Manual do instrumento de avaliação da atenção Primária à saúde : PCatool-Brasil – 2020 [recurso eletrônico] /
Ministério da saúde, secretaria de atenção Primária à saúde. – Brasília : Ministério da saúde, 237 p. : il. 2020
18. Lundberg SM, Lee SI. A unified approach to interpreting model predictions. Adv Neural Inf Process Syst. 2017;30:4766–75
19. Ferrer AP, Brentani AV, Sucupira AC, Navega AC, Cerqueira ES, Grisi SJ. The effects of a people-centred model on longitudinality of care and utilization pattern of healthcare services--Brazilian evidence. Health Policy Plan. 2014;29 Suppl 2(Suppl 2):ii107-13.doi:10.1093/heapol/czu077
20. Kessler M, de Lima S, Weiller T, Lopes L, Ferraz L, Eberhardt T, et al. Longitudinality of Primary Health Care: an evaluation from the perspective of users. Acta Paul Enferm. 2019;32(2):186-93
21. Prates ML, Machado JC, Silva LSD, Avelar PS, Prates LL, Mendonca ET, et al. Performance of primary health care according to PCATool instrument: a systematic review. Cien Saude Colet. 2017;22(6):1881-93.doi:10.1590/1413-81232017226.14282016
22. Stephens JH, Northcott C, Poirier BF, Lewis T. Consumer opinion on the use of machine learning in healthcare settings: A qualitative systematic review. Digit Health. 2025;11:20552076241288631.doi:10.1177/20552076241288631
23. Xu Z, Scharp D, Hobensack M, Ye J, Zou J, Ding S, et al. Machine learning-based infection diagnostic and prognostic models in post-acute care settings: a systematic review. J Am Med Inform Assoc. 2025;32(1):241-52.doi:10.1093/jamia/ocae278
24. Ravindranath R, Stein JD, Hernandez-Boussard T, Fisher AC, Wang SY, Consortium S. The Impact of Race, Ethnicity, and Sex on Fairness in Artificial Intelligence for Glaucoma Prediction Models. Ophthalmol Sci. 2025;5(1):100596.doi:10.1016/j.xops.2024.100596
25. Li Z, Zhou H, Xu Z, Ma Q. Machine learning and public health policy evaluation: research dynamics and prospects for challenges. Front Public Health. 2025;13:1502599.doi:10.3389/fpubh.2025.1502599
26. Wang ML, Bertrand KA. AI for all: bridging data gaps in machine learning and health. Transl Behav Med. 2025;15(1).doi:10.1093/tbm/ibae075
27. Smith LA, Oakden-Rayner L, Bird A, Zeng M, To MS, Mukherjee S, et al. Machine learning and deep learning predictive models for long-term prognosis in patients with chronic obstructive pulmonary disease: a systematic review and meta-analysis. Lancet Digit Health. 2023;5(12):e872-e81.doi:10.1016/S2589-7500(23)00177-2
28. Pires ALC, Costa FDS, D'Avila OP, Carvalho RV, Conde MCM, Correa MB, et al. Contextual inequalities in specialized dental public health care in Brazil. Braz Oral Res. 2024;38:e023.doi:10.1590/1807-3107bor-2024.vol38.0023
29. Costa FDS, Possebom Dos Santos L, Chisini LA. Inequalities in the use of dental services by people with and without disabilities in Brazil: a National Health Survey. Clin Oral Investig. 2024;28(10):540.doi:10.1007/s00784-024-05917-7
30. Leiva-Araos A, Contreras C, Kaushal H, Prodanoff Z. Predictive Optimization of Patient No-Show Management in Primary Healthcare Using Machine Learning. J Med Syst. 2025;49(1):7.doi:10.1007/s10916-025-02143-w
31. Wu Y, Jia M, Fang Y, Duangthip D, Chu CH, Gao SS. Use machine learning to predict treatment outcome of early childhood caries. BMC Oral Health. 2025;25(1):389.doi:10.1186/s12903-025-05768-y
32. Moannaei M, Jadidian F, Doustmohammadi T, Kiapasha AM, Bayani R, Rahmani M, et al. Performance and limitation of machine learning algorithms for diabetic retinopathy screening and its application in health management: a meta-analysis. Biomed Eng Online. 2025;24(1):34.doi:10.1186/s12938-025-01336-1
33. Zhou SN, Jv DW, Meng XF, Zhang JJ, Liu C, Wu ZY, et al. Feasibility of machine learning-based modeling and prediction using multiple centers data to assess intrahepatic cholangiocarcinoma outcomes. Ann Med. 2023;55(1):215-23.doi:10.1080/07853890.2022.2160008
34. Chen D, Liang S, Chen J, Li K, Mi H. Machine learning-based overall and cancer-specific survival prediction of M0 penile squamous cell carcinoma?A population-based retrospective study. Heliyon. 2024;10(1):e23442.doi:10.1016/j.heliyon.2023.e23442
35. Ashfaq A, Gray GM, Carapelluci J, Amankwah EK, Rehman M, Puchalski M, et al. Survival analysis for pediatric heart transplant patients using a novel machine learning algorithm: A UNOS analysis. J Heart Lung Transplant. 2023;42(10):1341-8.doi:10.1016/j.healun.2023.06.006
36. Yang J, Soltan AAS, Clifton DA. Machine learning generalizability across healthcare settings: insights from multi-site COVID-19 screening. NPJ Digit Med. 2022;5(1):69.doi:10.1038/s41746-022-00614-9
37. Chisini LA, Araujo CF, Delpino FM, Figueiredo LM, Filho A, Schuch HS, et al. Dental services use prediction among adults in Southern Brazil: A gender and racial fairness-oriented machine learning approach. J Dent. 2025;161:105929.doi:10.1016/j.jdent.2025.105929