Outcome-specific value of longitudinal clinical information and model complexity for dynamic prediction of type 2 diabetes-related outcomes: a pooled landmarking study.
Fu M, Liu X, et al. • Frontiers in endocrinology • 2026
The incremental value of longitudinal clinical information and model complexity for dynamic prediction of type 2 diabetes-related outcomes was outcome specific, with longitudinal history showing greater benefit for hypertension and MASLD, whereas algorithm differences were modest and outcome dependent.
Key Findings
Results
Pooled out-of-fold AUC ranged from 0.722 to 0.952 across the five representative models for predicting five type 2 diabetes-related outcomes.
646 patients with complete annual records from baseline through year 3 were included.
Landmark times at years 1 and 2 were used to predict the first recorded occurrence of each outcome during the subsequent year.
Internal validation used patient-level repeated nested cross-validation.
Model stability was evaluated using 500 patient-level bootstrap replicates with out-of-bag evaluation.
Results
Longitudinal clinical history improved prediction of hypertension for both elastic-net logistic regression and LightGBM.
Elastic-net showed a paired history-current ΔAUC of 0.097 (95% CI: 0.057 to 0.136) for hypertension.
LightGBM showed a paired history-current ΔAUC of 0.118 (95% CI: 0.081 to 0.158) for hypertension.
Both confidence intervals excluded zero, indicating a statistically meaningful improvement.
The primary estimand was a paired within-algorithm contrast with risk set, validation splits, preprocessing, tuning, calibration, and evaluation records held constant.
Results
Longitudinal clinical history improved prediction of MASLD for both elastic-net logistic regression and LightGBM, though the signal was smaller.
Elastic-net showed a paired history-current ΔAUC of 0.022 (95% CI: 0.002 to 0.042) for MASLD.
LightGBM showed a paired history-current ΔAUC of 0.025 (95% CI: 0.009 to 0.041) for MASLD.
The authors noted the smaller MASLD signal was supported primarily by the pooled L1-L2 analysis.
Both confidence intervals were narrow and barely excluded zero, suggesting a modest but detectable benefit.
Results
Adding longitudinal clinical history did not produce comparable improvements for coronary heart disease, diabetic kidney disease, or diabetic retinopathy.
The three outcomes (CHD, DKD, DR) showed no meaningful paired history-current AUC improvement for either algorithm.
This suggests the incremental value of longitudinal information is outcome specific rather than universal.
The comparison was conducted under identical risk sets, validation splits, preprocessing, tuning, calibration, and evaluation conditions across outcomes.
Results
Differences in predictive performance between elastic-net logistic regression and LightGBM were modest and outcome dependent.
Four progressively expanded feature sets were compared using both elastic-net logistic regression and Light Gradient Boosting Machine (LightGBM).
Neither algorithm consistently outperformed the other across all five outcomes.
The authors describe algorithm differences as 'modest and outcome dependent.'
This finding suggests that model complexity does not uniformly translate to improved prediction in this setting.
Methods
The study design used a pooled landmarking framework with patient-level repeated nested cross-validation to evaluate dynamic prediction models.
Landmark times at years 1 and 2 were pooled to increase effective sample size for model development and validation.
The primary history-current estimand held the risk set, validation splits, preprocessing, tuning, calibration, and evaluation records constant to isolate the effect of adding longitudinal history.
Model selection stability was evaluated using 500 patient-level bootstrap replicates with out-of-bag evaluation.
The cohort was single-center with complete annual T0-T3 follow-up for all 646 included patients.
Conclusions
The authors concluded that independent external validation is required before clinical application of the developed models.
The study was conducted in a single-center cohort, limiting generalizability.
The authors explicitly state 'Independent external validation is required before clinical application.'
The smaller MASLD signal was noted as being supported primarily by the pooled L1-L2 (elastic-net) analysis, warranting further confirmation.
Complete annual follow-up was required for inclusion, which may introduce selection bias not representative of routine clinical populations.
What This Means
This research suggests that using a patient's clinical history over time—rather than just their most recent measurements—can improve predictions of future diabetes complications, but only for some outcomes. The study followed 646 people with type 2 diabetes over three years, then used statistical models to predict whether each person would develop hypertension, heart disease, kidney disease, eye disease, or a liver condition in the following year. Researchers compared models that only used the most recent clinical data to models that incorporated the patient's full clinical history, and also compared a simpler statistical method to a more complex machine learning approach.
The key finding is that adding historical clinical information meaningfully improved predictions for hypertension and for a liver condition called MASLD (metabolic dysfunction-associated steatotic liver disease), but did not noticeably help for heart disease, kidney disease, or eye disease. The improvement for hypertension was substantial (roughly 10% better discrimination), while the improvement for MASLD was smaller but still detectable. Interestingly, the more complex machine learning algorithm (LightGBM) did not consistently outperform the simpler elastic-net logistic regression—differences between the two methods were small and varied by outcome.
This research suggests that the decision to collect and use longitudinal patient history in clinical prediction models should be made on an outcome-by-outcome basis rather than assumed to be universally beneficial. It also suggests that simpler statistical models may perform comparably to more complex machine learning approaches in some clinical contexts. However, because the study was conducted at a single center with a select group of patients who all had complete follow-up data, these findings will need to be confirmed in independent datasets before the models could be used in real clinical settings.
Fu M, Liu X, Shi T, Lu Y, Zou J, Huang X, et al.. (2026). Outcome-specific value of longitudinal clinical information and model complexity for dynamic prediction of type 2 diabetes-related outcomes: a pooled landmarking study.. Frontiers in endocrinology. https://doi.org/10.3389/fendo.2026.1964592