AI-generated summary
The authors report a retrospective cohort of 300 periodontitis patients from one center, comparing scaling and root planing with and without adjunctive low-level laser therapy. They developed a machine learning model called CaTGO to predict 6-month pocket depth and attachment outcomes and estimate individual treatment effects. The model showed high predictive accuracy on a validation set, but the study is retrospective, single-center, and the authors describe it as a predictive tool rather than proof of benefit.
- Participants
- 300
- Who or what was studied
- Patients with periodontitis treated at a single center, 150 receiving adjunctive low-level laser therapy and 150 receiving scaling and root planing alone
- Limits stated in the abstract
- Retrospective, single-center design with a model validated only on a 30 percent hold-out split from the same cohort, so external generalizability is not established.
Written by an AI model from this paper’s abstract only and checked automatically against it, but it can still be wrong. It is not medical advice. Read the full paper. See the editorial policy.
Abstract
Background
Adjunctive low-level laser therapy (LLLT) is proposed to improve periodontal healing post-SRP, but results are inconclusive and mostly reported as group averages. There's a need for decision tools to identify which patients, teeth, or sites benefit most from LLLT. We developed the Causal Tooth-Graph ODE Transformer (CaTGO), a "causal digital twin" model, to predict outcomes at patient, tooth, and site levels after periodontal therapy and estimate the individual treatment effect of LLLT.
Methods
This retrospective cohort study included 300 patients with periodontitis from a single center (150 received adjunctive LLLT and 150 received SRP alone). We recorded baseline pocket depth (PD), clinical attachment level (CAL), and patient factors (age, gender, and diabetes status) for treated patients, along with LLLT parameters. The CaTGO model uses a graph neural network for dental arch topology, neural ODEs for healing dynamics, a "Dose2Vec" embedding for LLLT doses, and a causal inference module to adjust confounding factors. It was trained (70% training, 30% validation) with the Adam optimizer (learning rate 0.001) and early stopping, and compared to baseline models.
Results
The CaTGO model achieved high predictive accuracy for 6-month outcomes (PD and CAL), with validation R2 values of 0.901 for PD and 0.880 for CAL, along with root-mean-square errors of 0.48 mm and 0.53 mm, respectively. It outperformed all tested models (ridge regression, random forest, gradient boosting) with a combined R2 of about 0.88. Predicted vs. actual outcomes had excellent correlation (Pearson r ≈ 0.95) and no significant residual bias, indicating good calibration.
Conclusions
The CaTGO digital twin predicted periodontal healing and identified patient-specific LLLT benefits, showing how graph-based deep learning and causal modeling can personalize therapy, guide clinicians, and improve decision-making.
Abstract reproduced under the CC BY license. Source: PubMed, a service of the National Library of Medicine.
How to read a single study: what to check before relying on this kind of research.
Terms in this paper: Low-level laser therapy, Laser, Dose. Each links to its definition in the glossary.
A summary of published research, not medical advice. One study rarely settles a question. See the disclaimer and how to read research.