Summary: In a machine-learning analysis of a periodontal-therapy trial in type 2 diabetes, a random forest model predicted treatment response with good accuracy, with lower baseline pocket depth, attachment loss, plaque, and HbA1c the main features favouring a good outcome.
PICO Summary
| Element | Detail |
|---|---|
| Population | 75 patients with periodontitis and type 2 diabetes; post-hoc analysis of a randomised trial, with 6-month follow-up. |
| Intervention | Machine-learning modelling (seven algorithms) of baseline periodontal, metabolic, and demographic factors to predict therapy response. |
| Comparison | Comparison of algorithms; leave-one-out cross-validation. |
| Outcome | Random forest performed best (AUC 0.83; accuracy 80%, sensitivity 64%, specificity 87%). Probing depth and clinical attachment level were the most important predictors; lower baseline pocket depth, attachment loss, plaque index, and HbA1c increased the chance of reaching the treatment endpoint. |
Expert Commentary
This is a methodologically modern take on a clinically grounded question, and the headline message is reassuringly intuitive rather than surprising: patients who start with less severe periodontal disease and better-controlled diabetes are the ones most likely to respond well to periodontal therapy. That probing depth and attachment loss dominated the model, with lower baseline HbA1c also favourable, simply quantifies what experienced clinicians already sense, and the bidirectional diabetes-periodontitis link gives it a coherent rationale. My caveats are squarely about the method and scale. A random forest trained on just 75 patients is highly prone to overfitting, an AUC of 0.83 from leave-one-out cross-validation in such a small sample is encouraging but fragile, and as a post-hoc analysis it generates rather than confirms predictors. External validation in a larger, independent cohort is essential before this becomes a usable tool. Can I use this with my patients? Conceptually, not computationally. It reinforces two things I can act on now, treating periodontal disease earlier before it becomes severe, and optimising glycaemic control to improve the odds that periodontal therapy succeeds, rather than relying on any specific predictive algorithm.
References
Castro Dos Santos N, Mangussi A, Ribeiro T, et al. Factors influencing the response to periodontal therapy in patients with diabetes: post hoc analysis of a randomized clinical trial using machine learning. J Appl Oral Sci. 2025;33:e20250211. doi:10.1590/1678-7757-2025-0211