Evaluation Of An Ai Iol Power Calculation Formula With A Novel Tabfpn Small-Sample Table Prior Model

To explore the construction of an intraocular lens (IOL) power calculation formula based on a novel Table Prior Data Fitted Network (TabPFN), evaluate its predictive accuracy for postoperative refractive error in cataract surgery, and provide a new method for accurate calculation of IOL power in clinical practice

retrospective cohort sutdy

We collected clinical data of 583 cataract patients (835 eyes) at Wuhan University Aier Eye Hospital from Jan to Oct 2025, divided into non-long AL (AL<26.0 mm, 618 eyes) and long AL (217 eyes) groups. Ocular biological parameters, IOL power and subjective refraction results ≥4 weeks postoperatively were extracted. A novel AI-based IOL power formula was built by combining TabPFN 2.5 model with thin lens vergence formula. Its predictive performance was compared with 6 classic IOL formulas (SRK-T, Cook-K6, Barrett Universal Ⅱ, EVO 2.0, Kane, Pearl DGS) and four control models (XGBoost, SVR, MLP, TabM). MAE, MedAE, and PE within ±0.50 D served as primary indicators, with statistical methods applied for comprehensive performance analysis.

Absolute prediction errors differed significantly across all AL subgroups and the full cohort (all P≤0.001). TabPFN 2.5 yielded the lowest MAE and MedAE in non-long AL (0.2995), long AL (0.3388) and full cohorts (0.3096), with the highest prediction rates in all error ranges; its ±0.5 D accuracy reached 74.4% in the long AL group, 9.3% higher than Barrett Universal II. The model showed minimal prediction deviation and was least affected by abnormal anatomy with variations in AL, Km, ACD and LT. Ablation experiments identified AL, Km, ACD and LT as effective predictors, while WTW, CCT, gender, eye side and age had no significant effect on model performance.

The new AI IOL power calculation formula based on TabPFN 2.5 can accurately capture high-order nonlinear relationships between eye parameters in small sample table mode, and exhibits better prediction accuracy and stability in both non long eye axis and long eye axis populations.