Enhancing Non-Surgical Root Canal Treatment Prognoses with Machine Learning

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Aug 25 2023

In a paper published in the journal Diagnostics, researchers examined the utilization of machine learning (ML) for improving prognostic accuracy in non-surgical root canal treatments (NSRCT). Traditional prognostic approaches, rooted in clinical experience, were investigated for enhancement through ML models such as Random Forest (RF) and K Nearest Neighbours (KNN). The results highlighted a noteworthy enhancement in sensitivity and accuracy in predicting NSRCT outcomes compared to conventional dentist-driven prognoses. The study underscores the potential for subsequent trials to evaluate the clinical utility of ML as an adjunctive tool in NSRCT prognostication.

Background

Apical Periodontitis (AP) emerges due to inflammation triggered by microorganisms within a tooth's root canal, necessitating timely intervention to avert tooth loss. With AP affecting nearly half of adults worldwide, NSRCT serves as the primary remedy, preserving tooth function and curbing infections. Nevertheless, the efficacy of NSRCT diminishes, encountering failure rates exceeding 15% at 5 years and 40% at 20 years. Clinicians rely on judgment to gauge treatment outlook for NSRCT decisions, introducing potential errors.

Contemporary dentistry capitalizes on extensive patient data harnessed through software tools to glean valuable insights. This study probes whether dentists conducting frequent NSRCT procedures can harness the database of the clinic and ML algorithms as a secondary opinion to heighten prognosis precision.

Related work

Previous research has extensively explored the relationship between treatment outcomes and risk factors in NSRCT. These risk factors span patient-related preoperative elements (e.g., pain, systemic conditions, tooth type, lesion size) and technical considerations like endodontic methods and materials. However, establishing associations between risk factors and outcomes marks just the initial phase; causation is not implied, necessitating testing for the predictive value of these factors.

While several studies have employed ML algorithms to predict endodontic outcomes, certain factors significantly linked to outcomes have exhibited limited predictive capacity. ML models incorporating all associated variables might perform suboptimally compared to judiciously chosen subsets. This underscores the significance of variable selection procedures to enhance the performance of ML models.

Proposed method

In this present paper, a retrospective analysis was carried out on patient case histories of AP who underwent NSRCT. The patient data were extracted from a private clinic in Mallorca, Spain, encompassing those set for dental appointments or facing emergencies. Selection criteria included patients without systemic diseases, receiving initial NSRCT, and possessing detailed records containing specific diagnostic and follow-up data.

The intervention involved 119 patients diagnosed with AP who underwent standardized NSRCT executed by the same endodontic specialist. The treatment protocol encompassed local anesthesia, rubber dam isolation, and comprehensive canal preparation and obturation using consistent materials. The treatment outcomes were assessed over at least nine years, focusing on clinical recovery measures and comparison of diagnostic radiographs to determine lesion regression.

Variables from eight domains, as outlined in a DCT, were meticulously recorded for each patient. The study defined treatment success as the absence of symptoms and symptoms warranting additional intervention, coupled with the disappearance of the lesion after NSRCT. Conversely, treatment failure occurred when either clinical or radiographic outcomes were unsatisfactory.

Statistical analyses and ML models were employed, utilizing the R software. Logistic Regression (LR), RF, Naive-Bayes (NB), and KNN models were evaluated. The Pearson chi-square test or Fisher exact test identified variables associated with outcomes. Subsequently, ML models were iteratively employed employing Leave-One-Out cross-validation (LOOCV) in conjunction with Backward Stepwise Selection (BSS) for variable selection. LOOCV allowed unbiased performance assessment, while BSS aided in identifying the most relevant features for predictive models. This iterative process led to the creation of ML models with specific subsets of associated variables. Performance metrics, including Sensitivity, Specificity, Positive Predictive Value, Negative Predictive Value, and Accuracy, were computed to gauge and compare model effectiveness.

Experimental analysis

The study employed retrospective analysis to investigate associations between various categorical variables and treatment outcomes in AP cases treated with NSRCT. Chi-square and Fisher’s exact tests identified nine variables linked to the outcome among 38 in the DCT. The minimal correlation was found through Spearman analysis, ensuring data validity. These associated variables encompassed demographic aspects like "Age" and "Highest Level of Education," patient-related factors such as "Smoking" and "Patient Cooperation," and clinical indicators like "Pain Relieved by," "Time-lasting of the Pain," "Periapical," and the dentist's "Estimated Prognosis," all contributing moderately to the outcome.

For outcome prediction, these associated variables were employed in a LOOCV + BSS procedure with four ML models: LR, RF, NB, and KNN. RF emerged as the best-performing model, surpassing others in predictive accuracy. While differences among ML models were not statistically significant based on confidence intervals, RF and KNN showed notable improvements in Sensitivity, Negative Predictive Value (NPV), and Accuracy compared to the dentist's prognosis. This highlights the potential of RF and KNN models in enhancing prognostic accuracy in NSRCT.

Conclusion

To sum up, the rapid rise of AI applications is propelled by factors such as increased data availability, advanced hardware, and user-friendly software. Private investments in AI have surged, particularly in data management and medical domains. In dentistry, AI has revolutionized diagnosis, treatment planning, and patient education. This approach explores the possibility of ML algorithms serving as a supplementary prognosis tool for NSRCT. By analyzing a dentist's assessment, the research uncovers associations between treatment outcomes and variables like "Age" and "Education Level." ML models, particularly RF, outperformed the prognosis of dentists in specific metrics, showcasing their potential to optimize NSRCT outcomes in real-world clinical settings. This paper demonstrated both efficiency and feasibility, laying a promising path for future research and practical implementation.