Managing Dental Appointment No-Shows: A Systematic Review of Machine Learning Applications

Abstract

Background

Missed appointments, commonly referred to as no-shows, are a persistent operational challenge in dental clinics, resulting in wasted clinical resources, disturbed workflows, and substantial financial losses. Machine learning, a predictive approach that identifies complex data patterns, has emerged as a promising tool for predicting patients at risk of missing scheduled dental visits. This systematic review aimed to synthesize and critically appraise machine learning-based prediction models developed to predict dental appointment no-shows.

Methodology

A comprehensive search of PubMed, Scopus, ScienceDirect, Google Scholar, and other databases was conducted, and eligible studies were screened using predefined inclusion and exclusion criteria. Data extraction and quality assessment were performed using the prediction model risk of bias assessment tool (PROBAST).

Results

The findings highlighted the applicability of machine learning models for no-show prediction with respect to dental appointments. However, direct comparisons across consistent datasets and evaluation metrics were limited. Key predictors included long lead times between the booking and appointment dates, missed confirmation messages, past no-show history, and temporal factors such as day of the week. Despite the reported performance of the models, existing studies were limited by small sample sizes, single-center designs, and a lack of external validation.

Conclusion

Overall, machine learning represents a valuable strategy for improving appointment management in dental clinics, but larger multicenter research and real-world implementation studies are required.

Introduction

Missed appointments lead to several challenges in the healthcare system. An appointment is considered “missed” when patients do not attend their pre-scheduled appointment or cancel within 24 hours of the scheduled appointment.¹ It has been estimated that 23% of scheduled appointments remain unattended on average.² Most importantly, no-shows result in the misuse of resources, an increase in wait times, and significant financial loss.¹

Appointments are scheduled in advance to allow time for equipment and administrative preparation according to each patient’s specific requirements. This means that for each appointment, healthcare staff prepare the examination rooms, equipment, and medical supplies, which remain unused despite being reserved and prepared in advance.³

No-show behaviour can cause significant financial and revenue losses for the healthcare system. The combination of wasted clinical time, unused staff hours, and administrative overhead results in a huge financial burden.^4,5 It was found that each missed appointment incurred a loss of approximately $196 per patient and a total of US$150 billion a year.⁶

Many solutions have been developed to treat this issue, including reminder calls, SMS messages, and emails.⁷ Additionally, statistical tools have been implemented to reduce the effects of missed appointments.^8,9 Despite these techniques and tools, this issue persists in several healthcare settings.

In recent years, machine learning (ML) models have rapidly emerged in the healthcare field, aiming to enhance the quality of the healthcare a patient receives. These models consist of statistical algorithms that automatically learn relationships from a given set of data and produce a fast, data-driven prediction. In healthcare, these models rely on electronic health records to improve clinical decision-making.¹⁰

Prediction of no-shows using ML models has been documented in the literature across various medical specialties. The results have shown that identifying individuals who are likely to miss appointments can help in scheduling strategies, minimizing wasted chair time, and reducing resource inefficiency. Although dental services typically experience a higher patient volume compared to many other medical fields, to the best of our knowledge, evidence on no-show prediction in dental clinics remains limited.^2,11

Therefore, a systematic literature review was needed to merge current evidence, identify common predictors, evaluate the performance of existing models, and highlight gaps for future research aimed at improving scheduling efficiency, reducing resource, and financial waste in dental care.

In this review, we summarized the predictive performance and validation approaches used in models developed to predict no-show events. In addition, it sought to describe the methodological characteristics of the included studies, including the types of models applied, the predictors incorporated, and the data sources utilized. Lastly, the study evaluated whether the financial losses associated with predicted no-show events have been reported in the existing literature.

Methodology

This systematic review was conducted using the PRISMA guidelines, followed by predefined inclusion and exclusion criteria. Additionally, this protocol was registered in PROSPERO under the ID number CRD420261302026.

Eligibility criteria

The inclusion criteria were as follows: empirical studies that developed no-show prediction models and used hospital electronic records as a data source, studies set in a dental clinic, and patients with scheduled dental visits. The primary dependent variable in these studies was no-show (i.e., absence from scheduled appointments), and the studies were required to report model performance metrics, key predictive features, and implementation in real-world settings. Publications in English from 2010 to 2025 were included. No restrictions regarding publication type or grey literature were applied during the search process, and all eligible studies were considered for inclusion if accessible.

On the other hand, the exclusion criteria were studies that provided only descriptive attendance analysis without a predictive model, lacked transparent data access to ensure reliability, or failed to report appropriate ethical approval.

Data extraction

This systematic review followed the PICOS framework to define eligibility criteria and guide data extraction.

Population (P): Patients who booked appointments in a dental clinic. Information such as the patient’s age, country, clinical type, sample size, hospital settings, and rate of no-shows were extracted.

Index model (I): Any type of machine learning model used for predicting appointment no-show events, with the exclusion of descriptive-only studies.

Comparison (C): Any available comparison between different machine learning algorithms.

Outcomes (O): The extracted outcomes included model performance measures, such as the area under the receiver operating characteristic curve (AUC), accuracy, F1-score, precision, recall, specificity, and sensitivity.

Study design (S): This determined whether a study was observational or retrospective, applying machine learning methods to healthcare appointment data.

The databases searched included PubMed, Scopus, Web of Science, IEEE Xplore, ScienceDirect, and Google Scholar. Data extracted from each eligible study included bibliographic information (including author, year, country, and journal), predictors, or features used, assessment of the model assumptions, hyperparameter tuning strategies, feature importance methods (including SHAP), handling of missing data and class imbalance, and internal validation techniques, such as train/test split or k-fold cross-validation. Finally, any reported limitations were listed by the authors. The database search was conducted on October 11, 2025, and continued to November 16, 2025.

Risk of bias and quality assessment

For each included study, the quality of the papers was determined using the prediction model risk of bias assessment tool (PROBAST). The tool evaluated 4 domains – participants, predictors, outcomes, and analysis. Each study was rated as having a low, high, or unclear risk of bias according to published guidance.¹² Data extraction, title and abstract screening, and full-text assessment were conducted independently by 2 reviewers (SK and MB), who were blinded to each other’s decisions during the screening process. Any disagreements were resolved through discussion and, when necessary, consultation with the third reviewer (AA).

Keywords for search strategy

The search strategy combined medical subject headings (MeSH) and free-text terms to ensure a comprehensive and sensitive search. MeSH terms included ‘Dentistry’, ‘Hospitals’, ‘Outpatient Clinics, Hospital’, ‘Artificial Intelligence’, and ‘Machine Learning’. Free-text terms included dental clinic, dental appointment, no-show, missed appointment, non-attendance, and prediction. These terms were combined using Boolean operators and adapted to the syntax of each database, as seen in Table 2 in Supplementary Material 1.

Table 2.

Machine learning models methods for included studies.

Author	Predictors	Model validation	Missing data handling	Tuning	Class imbalance handling
Almutairi, T. et al.13	Patients: Age, sex, occupation, and no-show ratios Physicians: specialty, gender, and credentials Appointment: Date, time, type, booking channel, and confirmation record Financial data: Coverage, affordable fee percentage, and self-payment for the appointment	Cross Validation: k-fold = 5	Yes	Yes	Yes
Alabdulkarim,Y. et al.14	Patients: Age, marital status, gender, and nationality. Appointments: Date, time, booking date and time, appointment duration, confirmation SMS, seasonality	Simple Split: 90%/10%	Yes	Yes	No
Alqahtani, H. et al.15	Patient: Location, age, gender Appointments: Weeks, time, day	Not Mentioned	Not Mentioned	No	Not Mentioned
Cuevas-Nunez, M. et al16	Patients: Age, gender, presence, comorbidities translation required, presence or absence of insurance Appointment: month, year, day, time, Duration of appointment, confirmation status, lead time	Simple Split: 70%/30%	Not Mentioned	No	Not Mentioned

Results

Selection of eligible studies

A total of 85 published studies were identified, out of which 9 were removed for duplication and 68 for not applying ML models or using any type of predictive model, indicating that they were purely descriptive studies. Therefore, eight studies were included for full-manuscript assessment. However, the reviewers were unable to retrieve the full manuscript of one conference paper. Thus, seven studies underwent eligibility assessment. After eligibility assessment, a total of 3 studies were excluded out of seven due to a focus on disease prediction rather than no-show prediction, the use of open-access data, and inability to obtain the full text despite attempts to retrieve it through institutional and collaborating library subscriptions, as seen in Figure 1.

Fig. 1.

PRISMA diagram for search flow.

Prediction model risk of bias assessment of the included studies

After assessing the quality of the 4 included studies using PROBAST, most were judged to have a high risk of bias, mainly due to limitations in the analysis domain, such as a lack of calibration assessment, limited external validation, and insufficient handling of class imbalance, which resulted in modest model performance. In addition, incomplete reporting of outcomes and narrow participant selection contributed to unclear risk in some domains. These findings highlight the need for more carefully designed and transparently reported studies on prediction models in this field, as seen in Supplementary Material 1.

Characteristics of the population

The 4 studies were published between 2022 and 2024 and were conducted in 2 countries – Saudi Arabia and Canada. A total of 1,162,030 patients were included in this systematic review. Two of the included studies recruited adult patients with an average age older than 20 or 29 years. Only one study included adolescents aged 12 years, and one study did not report information on age.

A multicenter study design was implemented in one study, which included nine dental clinics in one region of Saudi Arabia. All studies were retrospective, with no-show rates ranging from 13.5% to 48%, as seen in Table 1.

Table 1.

Characteristics of the included studies.

Author	Year of publication	Name of journal	Population
			City, Country	Setting	Age	Sample size	No-Show Rate
Almutairi, T. et al.13	2024	Informatics in Medicine Unlocked	Dammam, Saudi Arabia	Nine Dental Clinic	From 12 to 35	260,951	Not mentioned
Alabdulkarim,Y. et al.14	2022	PeerJ Computer Science	Riyadh, Saudi Arabia	One Dental Clinic	Average: 29	196,018	48%
Alqahtani, H. et al.15	2023	Cureus	Riyadh, Saudi Arabia	One Dental Clinic	≥20	14,066	31%
Cuevas-Nunez, M. et al16	2023	Journal of Dental Education	Surrey, Canada	Dental School Clinic	NA	690,995	13.5%

Characteristics of machine learning models

Commonly selected features were age, gender, and demographics. Only one study considered patients’ occupations, and another included location. As for appointment-related variables, all studies included common factors, such as appointment time, day, week, lead time, and SMS confirmation. However, only one study accounted for seasonality in relation to no-shows, especially events in the Islamic calendar, such as Eid and Ramadan.

One of the 4 studies included finance-related variables, such as insurance coverage, affordable fee percentage, and self-payment for appointments. However, none of the included studies directly quantified the financial losses associated with no-show events. Additionally, only one study considered the health status of patients, as seen in Table 1.

In Table 2, most methodological approaches used to build the models in the included studies led to high risk of bias. Several studies did not report calibration assessment, external validation, or appropriate class imbalance handling. In addition, some studies lacked transparent reporting regarding missing data handling and model optimization or tuning strategies. These methodological limitations reduce confidence in the reported model performance and limit the generalizability and clinical applicability of the findings.

Outcome characteristics

Most of the included studies reported discrimination metrics, such as AUC and accuracy. However, calibration measures were rarely assessed. Ensemble models were commonly used for predicting no-shows to appointments across the included studies, with accuracy ranging from 76% to 81%. The most important predictors of no-show identified across the included studies were SMS confirmation status, prior history of no-show, and appointment timing. Patients who did not confirm their appointment via SMS were at a higher risk of non-attendance. Moreover, those with a previous history of no-show were more likely to miss future appointments, and appointments scheduled on Mondays or during the first week of the month were associated with increased no-show rates. Lastly, the longer the duration of time between requesting the appointment and the date of the actual appointment, the higher the risk of no-show among patients, as presented in Table 3.

Table 3.

Machine learning models outcome for included studies.

Author	Index Model (I)	Study design (S)	Period collection	Feature importance/SHAP	Performance outcome (O): discrimination and calibration	Limitation
Almutairi, T. et al.13	MLP, Decision Trees, Random Forest.	Retrospective	4 years	SHAP Day SMS confirmation Previous no-show	Accuracy: 81%, Precision: 81%, Recall: 93%, F1: 87%, AUC: 83% Calibration is not mentioned.	Real-application, Lack of variables
Alabdulkarim,Y. et al.14	Logistic Regression, Random Forest, Gradient Boosting.	Retrospective	One year	Feature importance Longer duration No-show history	F1: 0.661 AUC: 0.718. Calibration not mentioned.	Home address, Distance to the clinic, Daily weather.
Alqahtani, H. et al.15	CTree Random Forest.	Retrospective	19 weeks	Feature importance First week Mondays Female	No	Available data, Generalizability.
Cuevas-Nunez, M. et al16	Decision tree Bagging Regressor, Random Forest, Gradient Boosted AdaBoost Regression, XGBoost, Neural Network, Logistic Regression.	Retrospective	June 2013 to November 2022	Feature importance Confirmation Age, Number of visits before appointment, Previous no-show Lead time	F1: 0.41, Accuracy: 0.76. Calibration not mentioned	Cost-effectiveness, Generalizability.

Despite the reported model performance, a few limitations were mentioned in the included studies. First, the limited number of published studies on dental clinics and the reliance mainly on a single center or a single clinic limits the generalization of the studies.

Secondly, the available data, since they are extracted from the hospital database, rarely contain other factors, such as distance from the patient’s home to the hospital, and weather conditions. Finally, it is not yet known how these models may act in a real clinical setting, as there are no studies yet examining the application of scheduling systems or the cost benefits of these models, as shown in Table 3.

Discussion

Machine learning provides insights into patient behavior. This helps clinics identify individuals at a high risk of missing appointments and, thereby, implement timely interventions. One study mentioned how predictive analytics can support targeted reminders, flexible rescheduling, outreach calls, and optimized overbooking policies, which are strategies that have been shown to reduce missed visits and improve clinical efficiency.¹¹

The findings of the included studies showed that ensemble models were commonly used to predict no-shows to appointments. Previous research in various medical settings has demonstrated that ensemble models effectively capture complex patterns and address class imbalances, highlighting their applicability to real-world clinical scheduling.17, 18, 19

The findings of the included studies suggest that age is among the influential variables associated with the target outcome. Adults aged between 20 and 40 years presented a higher probability of skipping their appointments. On the other hand, older patients were more likely to attend their appointments. These findings align well with other descriptive studies that aimed to assess the no-show rate in multiple settings.

Younger patients exhibited higher no-show rates, especially for appointments scheduled with longer lead times.20, 21, 22 The literature provides an explanation for this relation, as younger patients tend to have work-related commitments that can lead to them missing their appointments.²¹ Additionally, younger patients often see attending their appointments as less urgent or optional.³ Finally, younger patients are likely to rely on unstable transportation, which can limit their likelihood of showing up for appointments.²² These reasons highlight the importance of examining patients’ occupations; however, only one study included in this systematic review considered occupation as a variable. Regarding transportation availability, none of the included studies incorporated it as a feature, although it could influence the direction of future studies.

Other important features identified across the 4 included dental ML studies, including SMS confirmation, long lead time, and day of the appointment, had an influence on the no-show rate. These findings are consistent with previous studies, where patients who did not confirm their attendance through the received SMS message showed a higher risk of not attending their appointments. This pattern is often considered an early sign of a low intention to attend.²³ Additionally, past no-show behavior tends to have a greater influence on the target outcome, showing its importance as a predictor when predicting the outcome. A high history of no-show is often seen as a repeated behaviour caused by ongoing barriers.³

Lastly, a longer lead time between booking and appointment date had a high risk of non-attendance, as patients could forget or lose interest when appointments are scheduled too far in advance.²⁴ Research suggests that attendance may vary depending on whether appointments fall on busy weekdays, weekends, or holidays. This shows the need to include seasonality related to both Islamic and non-Islamic holidays or public events when analysing and predicting appointment attendance.

Collectively, no-show behaviour may not always reflect patient preference alone, as social and healthcare access barriers can also affect attendance. Factors such as transportation, socioeconomic status, language barriers, occupation, and access to care may influence whether patients attend their appointments.

In real clinical settings, prediction models may support targeted reminder, flexible rescheduling, and optimized scheduling strategies. However, these models should be implemented carefully to avoid unfair overbooking or unintentionally increasing healthcare disparities among vulnerable patient populations.

Although ML has been increasingly applied in various areas of dental care, such as dental imaging, prosthodontics, and other contemporary dental AI applications, operational applications in dentistry, including no-show prediction, are still in their early stages.25, 26, 27, 28 The limited number of published studies, absence of calibration assessment, inconsistent handling of class imbalance, and limited external validation reduce confidence in the reported model performance. Therefore, these findings should be interpreted with caution despite the growing interest in AI applications in dental healthcare.

Future research should focus on improving the generalizability by including multiple settings from different regions locally and internationally. Additionally, including personalized factors such as distance, transportation, weather, occupation, and psychosocial or health-related factors may improve model accuracy. Moreover, real clinical application and validation are important and missing steps in the literature, as it will determine whether these models truly improve scheduling efficiency and quality of patient care. Many of the included studies focused mainly on discrimination measures, such as accuracy and AUC, while calibration was rarely reported. In addition, some studies did not adequately address class imbalances, which may have influenced their reported performance.

To this day, no prior study has predicted no-show and subsequently calculated the financial losses of the resulting no-shows, and external validation was limited, as most studies relied on internal validation methods only.

These limitations should be considered when interpreting the results of this review. The continuous evaluation of this global issue can reveal the true impact and whether the implementation of prevention measures, such as ML, has actually reduced this burden.

Conclusion

This review summarizes the current evidence on the use of ML-based models to predict appointment no-shows in dental clinics. The included studies showed the performance of ensemble models including random forest and gradient boosting, although comparisons with traditional statistical models were not always conducted under consistent evaluation frameworks. The models included common and important predictors, such as SMS confirmation status, previous no-show behaviour, appointment timing, and scheduling lead time, to improve model performance.

Overall, this systematic review adds to the growing body of literature in improving scheduling efficiency and resource allocation in dental clinics. However, the current evidence remains limited and is based on only 4 studies with a generally high risk of bias. External validation, calibration assessment, prospective testing, cost-effectiveness analysis, and implementation studies in real dental care settings are needed to better evaluate the clinical and operational applicability of these models.

Future research should focus on multicenter data sources and model validation on real scheduling systems to assess operational feasibility. The application of findings from such future studies can support appointment management strategies and enhance patient care delivery in dental healthcare environments. Finally, the financial impact of missed appointments has not yet been addressed. Including financial analysis in future prediction models would allow healthcare systems to quantify the economic burden of no-shows and evaluate whether preventive measures truly reduce this burden.

Ethical approval

Ethical approval was not required for this systematic review, as it is based on previously published studies. This review was conducted in accordance with the principles of the Declaration of Helsinki.

Consent to participate

Not applicable. Participant consent is not required for this study.

Consent for publication

Not applicable. This review does not include any individual person’s identifiable data.

Data availability

All data used in this systematic review were extracted from publicly available published studies. The search strategies, included studies, and extracted data were well described in the methodology section.

Conflict of interest

None disclosed.