Skip to content
Log in

Similar papers

Open access Jul 2026

An explainable machine learning framework for accurate prediction of postoperative anterior chamber depth in highly myopic cataract surgery

Purpose To develop and validate machine learning models for predicting postoperative anterior chamber depth (ACD) in highly myopic cataract patients based on preoperative biometric parameters. Methods This prospective study enrolled 203 eyes of 127 highly myopic patients who underwent phacoemulsification and intraocular lens (IOL) implantation between January 2024 and December 2025. Ocular biometric parameters were measured preoperatively and at 3 months postoperatively. A dual feature selection strategy combining Least Absolute Shrinkage and Selection Operator (LASSO) regression and Boruta algorithm was employed to identify important predictors of postoperative ACD in these samples. We compared five machine learning algorithms and evaluated their performance using the coefficient of determination (R 2), mean absolute error (MAE), root mean square error (RMSE), and accuracy within ±0.1 mm and ±0.2 mm. Subsequently, Shapley Additive Explanations (SHAP) method was applied to interpret the optimal model’s feature importance. Results Six predictors were identified for model construction: axial length ACD/lens thickness ratio (ACD/LT), white-to-white distance (WTW), ACD at 90° (ACD90), horizontal position angle, and ACD + LT/2. Among all models, Random Forest algorithm demonstrated the best predictive performance, achieving an R 2 of 0.8259, mean absolute error of 0.0604 mm, and root mean square error of 0.0722 mm in the test set. The accuracy within ±0.1 mm reached 80.49%, and within ±0.2 mm reached 100%. SHAP analysis revealed that ACD + LT/2 was the most important predictor, followed by WTW, ACD90, and horizontal position angle. Conclusion This study successfully developed and validated an explainable RF-based machine learning model for the prediction of postoperative ACD in highly myopic cataract patients, which supports more reliable IOL power calculation and offers a practical tool for optimizing surgical planning in highly myopic eyes.

Yuyang Yang, Hao Cui, Jiajia Gao et al. · 0 citations
Open access Jul 2026

The Quest to Predict Surgically Induced Astigmatism After Cataract Surgery: Lessons for Toric IOL Planning.

PURPOSE This study compared traditional statistical models with machine learning algorithms for predicting surgically induced astigmatism after cataract surgery, aiming to identify the most accurate and generalizable method among linear regression, regression trees, random forests, and neural networks. METHODS Retrospective analysis was performed on 321 eyes (321 patients) undergoing phacoemulsification at a tertiary center. Data were standardized and surgically induced astigmatism was modeled using three vector-based outcomes: KEQ.post (keratometric equivalent power), KAST0.post (horizontal astigmatism component), and KAST45.post (oblique component). Preoperative variables served as predictors. Data were split into training (60%) and test (40%) sets. Four predictive models were developed and evaluated on unseen data. Performance was assessed using mean squared prediction error, and variable importance analyses identified key predictors. RESULTS Linear regression achieved the best out-of-sample performance across all outcomes (e.g. KEQ.post mean squared prediction error = 0.043). Tree-based models performed slightly worse, while neural networks showed substantial overfitting with markedly higher test errors. Preoperative astigmatism and corneal radii were the strongest predictors. CONCLUSION KAST0 and KAST45 were not predictably modeled, whereas changes in KEQ.post were. Multivariable linear regression provided the most accurate and reliable predictions, while more complex machine-learning models (especially neural networks) overfit the limited dataset and offered no clinical advantage. The relationship between preoperative biometrics and postoperative equivalent power appears predominantly linear, though larger datasets may enhance machine-learning performance.

Amanda Pan, K. P. Kaiser, Stefan Raidl et al. · 0 citations
Open access Aug 2026

Role of surgeon seniority in predicting surgically induced astigmatism after phacoemulsification surgery: a machine learning study

To evaluate whether surgeon seniority influences postoperative astigmatic change after phacoemulsification cataract surgery and to assess the performance of machine-learning models in predicting postoperative astigmatic outcomes using preoperative clinical data. This study included 169 eyes of 169 patients who underwent phacoemulsification cataract surgery. Surgeons were categorized as residents ( n  = 101) or specialists ( n  = 68). Astigmatism was quantified using both classical magnitude-based analysis and vector decomposition. Potential determinants of postoperative astigmatic change were evaluated using multivariable regression analyses. Machine-learning models, including Gradient Boosting, K-Nearest Neighbors, Decision Tree, and Support Vector Machine models, were used to predict classical and vector astigmatism outcomes. Model performance was assessed using mean absolute error, root mean square error, and the coefficient of determination. Of the 169 eyes, 101 were operated on by residents and 68 by specialists. Postoperative astigmatic change did not differ significantly between resident- and specialist-performed procedures in either classical or vector analyses. Within the resident group, increasing surgical seniority was not associated with postoperative astigmatic change. In multivariable analysis, preoperative cylindrical error was the only significant predictor in the vector model, whereas no significant predictor was identified in the classical model. Overall, machine-learning models demonstrated limited predictive performance, with only modest utility for classical astigmatism and poor explanatory performance for vector outcomes. Surgeon seniority was not a significant determinant of surgically induced astigmatism after phacoemulsification cataract surgery. Machine-learning models based on preoperative clinical data provided only limited predictive value, particularly for vector astigmatism outcomes. Incorporating intraoperative variables may improve future predictive performance.

Denizcan Özizmirliler, C. Engin, Özlem Özkan et al. · 0 citations
Open access Jul 2026

Predicting gingival embrasure risk after invisible orthodontics using multimodal data and machine learning

Objective To develop and validate a risk prediction model for gingival embrasures after clear aligner therapy using multimodal oral data. Methods A retrospective study of 340 patients (December 2022–June 2025) was randomly divided into training (n = 238) and validation (n = 102) sets (7:3). Univariate analysis, multivariate logistic regression, and least absolute shrinkage and selection operator regression were applied to identify independent risk factors. Three machine learning models – random forest (RF), logistic regression, and support vector machine – were constructed based on seven core variables. Model performance was assessed using area under the receiver operating characteristic curve (AUC), calibration curves, decision curve analysis, and Shapley Additive Explanations (SHAP) values for interpretability. Results No significant baseline differences existed between sets (p > 0.05). Seven indicators were identified (p < 0.05). Multivariate analysis confirmed percentage of bleeding on probing-positive sites, interproximal alveolar bone height, and relative movement of adjacent teeth at target site as independent risk factors, while gingival thickness, proximal contact area, interdental papilla height, and buccal/lingual bone plate thickness were protective factors (p < 0.05). The RF model performed best: training AUC = 0.849 (95% CI: 0.786–0.912), validation AUC = 0.815 (95% CI: 0.720–0.910), with good calibration and net benefit. SHAP analysis highlighted gingival thickness and interproximal alveolar bone height as key predictors. Conclusion A risk prediction model for post-clear aligner gingival embrasures was successfully developed and validated using multimodal oral data, with RF as the optimal algorithm. The model exhibits good discrimination, calibration, and clinical utility, which can be used as an objective auxiliary tool for individualized risk prediction following clear aligner therapy and supplement traditional clinical empirical judgment.

Haiyan Wang, Hanfei Shi, Liping Fan et al. · 0 citations
Open access Jul 2026

Development of Machine Learning Models for Predicting Surgical Site Infection After Spinal Surgery

Background/Objectives: Surgical site infection (SSI) remains a clinically important complication after spinal surgery. This study developed and assessed machine learning approaches for predicting postoperative SSI using routinely collected preoperative clinical variables, with emphasis on calibration and clinical applicability. Methods: In this retrospective single-center study, four prediction models were developed in patients undergoing spinal surgery: logistic regression, random forest, gradient boosting, and XGBoost. Model training used five-fold stratified cross-validation, and performance was evaluated using a hold-out internal test set. Performance was assessed using the area under the receiver operating characteristic curve (AUC), area under the precision–recall curve (AUPRC), sensitivity, precision, F1 score, Brier score, and calibration slope. SHAP analysis was performed to evaluate model interpretability. Results: The incidence of SSI was 16.6%. In cross-validation, discrimination performance was broadly comparable across models, with logistic regression showing the highest observed AUC (0.814) and AUPRC (0.484). In the hold-out test set, the same model showed the highest AUC (AUC 0.806, 95% CI 0.757–0.852) and the highest sensitivity (0.758). Calibration performance varied across models. SHAP analysis identified C-reactive protein, hemoglobin, albumin, and white blood cell count as the most influential predictors. Perioperative variables provided only modest incremental predictive value. Conclusions: Machine learning models showed acceptable performance for predicting SSI after spinal surgery. Logistic regression demonstrated performance comparable to that of the evaluated machine learning models, suggesting that conventional statistical approaches may remain clinically useful in structured datasets. Preoperative clinical and laboratory variables were the major contributors to prediction, supporting their use for routine preoperative risk stratification.

Kwang-Ryeol Kim, Gi-Young Park, Dong Hyuck Kim et al. · 0 citations
Open access Jul 2026

Preoperative artificial intelligence-based risk model for surgical reintervention after microsurgical free flap reconstruction.

BACKGROUND Flap-related vascular complications requiring surgical reintervention remain a source of morbidity after microsurgical free flap reconstruction and preoperative risk estimation relies on clinical judgement. Therefore, we developed and internally validated a strictly preoperative multivariable model for this outcome. METHODS A retrospective cohort of 650 consecutive adults at two high-volume referral centres (649 analysable) was analysed. The primary outcome was unplanned re-exploration within 30 days for arterial, venous, mixed thrombosis, or clinically significant vasospasm. Nineteen preoperative predictors were considered; intraoperative and surgeon variables were excluded. Least absolute shrinkage and selection operator (LASSO), random forest, and XGBoost were fitted on a 70% training partition and evaluated on a 30% test set. Performance was assessed using discrimination (AUC), calibration (intercept, slope, ICI, and E/O), and Brier score. Reporting followed TRIPOD; risk of bias used the four-domain PROBAST. RESULTS Event rate was 17.1% (111/649). XGBoost achieved the highest discrimination (AUC 0.84; 95% CI 0.74-0.92) and relatively better calibration (intercept 0.02, slope 0.87, ICI 0.039, E/O 0.89). Random forest showed comparable discrimination (AUC 0.82; 0.71-0.90) but poorer calibration (slope 1.33). LASSO demonstrated the lowest discrimination (AUC 0.79; 0.69-0.88). Prior oncologic history, surgical indication, and flap composition were influential predictors. A decile table from XGBoost showed a monotonic gradient, with events rising from ≤10% in lower deciles to 80% (95% CI 58-92) in the top decile. CONCLUSIONS XGBoost is retained as the principal model based on combined superiority in discrimination and calibration, with random forest as a robust comparator. The model is not decision-ready, and prospective external validation with recalibration is required before clinical adoption. LAY SUMMARY Using 649 analysable free flap reconstructions, we developed preoperative models to predict unplanned surgical reintervention for flap-related vascular complications within 30 days. XGBoost performed the best (AUC 0.84), supporting risk stratification before surgery, but prospective external validation is needed before clinical use.

Luis Arturo Molina Laguna, A. Porras-Ramírez, Giovanni Montealegre et al. · 0 citations

Related blog posts

MIT News · Artificial Intelligence Aug 17, 2026

Q&A: Rethinking how innovation happens

In his latest book, Professor Eugene Fitzgerald examines the forces that turn breakthroughs into value — and why innovation resists simple formulas.