Comparing machine learning approaches to incorporate time-varying covariates in predicting cancer survival time

Drug exposure as a predictor in diabetic retinopathy risk prediction models; a systematic review and meta-analysis

PURPOSE

To conduct a systematic review to assess drug exposure handling in diabetic retinopathy (DR) risk prediction models, a network-meta-analysis to identify drugs associated with DR and a meta-analysis to determine which drugs contributed to enhanced model performance.

DESIGN

Systematic review and meta-analysis.

METHODS

We included studies presenting DR models incorporating drug exposure as a predictor. We searched EMBASE, MEDLINE and SCOPUS from inception to December 2023. We evaluated the quality of studies using the Prediction model Risk of Bias Assessment Tool and certainty using GRADE. We conducted network meta-analysis and meta-analysis to estimate the odds ratio (OR) and pooled C-statistic, respectively, and 95% confidence intervals (CI) (PROSPERO: CRD42022349764).

RESULTS

Of 5,653 records identified, we included 28 studies of 678,837 type 1 or 2 diabetes participants, of which 38,579 (5.7%) had DR. A total of 19, 3 and 7 studies were at high, unclear, and low risk of bias, respectively. Drugs included in models as predictors were: insulin (n=24), antihypertensives (n=5), oral antidiabetics (n=12), lipid-lowering drugs (n=7), antiplatelets (n=2). Drug exposure was modelled primarily as a categorical variable (n=23 studies). Two studies handled drug exposure as time-varying covariates, and one as a time-dependent covariate. Insulin was associated with an increased risk of DR (OR= 2.50; 95%-CI: 1.61-3.86). Models that included insulin (n=9) had a higher pooled C-statistic (C-statistic=0.84, CI: 0.80-0.88), compared to models (n=9) that incorporated a combination of drugs alongside insulin (C-statistic= 0.79, CI:0.74-0.84), as well as models (n=3) not including insulin (C-statistic =0.70, CI: 0.64-0.75). Limitations include the high risk of bias and significant heterogeneity in reviewed studies.

CONCLUSION

This is the first review assessing drug exposure handling in DR prediction models. Drug exposure was primarily modelled as a categorical variable, with insulin associated with improved model performance. However, due to suboptimal drug handling, associations between other drugs and model performance may have been overlooked. This review proposes the following for future DR prediction models: 1) evaluation of drug exposure as a variable, 2) use of time-varying methodologies, and 3) consideration of drug regimen details. Improving drug exposure handling could potentially unveil novel variables capable of significantly enhancing the predictive capability of prediction models.

Creation of a machine learning-based prognostic prediction model for various subtypes of laryngeal cancer

Depending on the source of the blastophore, there are various subtypes of laryngeal cancer, each with a unique metastatic risk and prognosis. The forecasting of their prognosis is a pressing issue that needs to be resolved. This study comprised 5953 patients with glottic carcinoma and 4465 individuals with non-glottic type (supraglottic and subglottic). Five clinicopathological characteristics of glottic and non-glottic carcinoma were screened using univariate and multivariate regression for CoxPH (Cox proportional hazards); for other models, 10 (glottic) and 11 (non-glottic) clinicopathological characteristics were selected using least absolute shrinkage and selection operator (LASSO) regression analysis, respectively; the corresponding survival models were established; and the best model was evaluated. We discovered that RSF (Random survival forest) was a superior model for both glottic and non-glottic carcinoma, with a projected concordance index (C-index) of 0.687 for glottic and 0.657 for non-glottic, respectively. The integrated Brier score (IBS) of their 1-year, 3-year, and 5-year time points is, respectively, 0.116, 0.182, 0.195 (glottic), and 0.130, 0.215, 0.220 (non-glottic), demonstrating the model's effective correction. We represented significant variables in a Shapley Additive Explanations (SHAP) plot. The two models are then combined to predict the prognosis for two distinct individuals, which has some effectiveness in predicting prognosis. For our investigation, we established separate models for glottic carcinoma and non-glottic carcinoma that were most effective at predicting survival. RSF is used to evaluate both glottic and non-glottic cancer, and it has a considerable impact on patient prognosis and risk factor prediction.

Dynamic prediction of malignant ventricular arrhythmias using neural networks in patients with an implantable cardioverter-defibrillator

BACKGROUND

Risk stratification for ventricular arrhythmias currently relies on static measurements that fail to adequately capture dynamic interactions between arrhythmic substrate and triggers over time. We trained and internally validated a dynamic machine learning (ML) model and neural network that extracted features from longitudinally collected electrocardiograms (ECG), and used these to predict the risk of malignant ventricular arrhythmias.

METHODS

A multicentre study in patients implanted with an implantable cardioverter-defibrillator (ICD) between 2007 and 2021 in two academic hospitals was performed. Variational autoencoders (VAEs), which combine neural networks with variational inference principles, and can learn patterns and structure in data without explicit labelling, were trained to encode the mean ECG waveforms from the limb leads into 16 variables. Supervised dynamic ML models using these latent ECG representations and clinical baseline information were trained to predict malignant ventricular arrhythmias treated by the ICD. Model performance was evaluated on a hold-out set, using time-dependent receiver operating characteristic (ROC) and calibration curves.

FINDINGS

2942 patients (61.7 ± 13.9 years, 25.5% female) were included, with a total of 32,129 ECG recordings during a mean follow-up of 43.9 ± 35.9 months. The mean time-varying area under the ROC curve for the dynamic model was 0.738 ± 0.07, compared to 0.639 ± 0.03 for a static (i.e. baseline-only model). Feature analyses indicated dynamic changes in latent ECG representations, particularly those affecting the T-wave morphology, were of highest importance for model predictions.

INTERPRETATION

Dynamic ML models and neural networks effectively leverage routinely collected longitudinal ECG recordings for personalised and updated predictions of malignant ventricular arrhythmias, outperforming static models.

FUNDING

This publication is part of the project DEEP RISK ICD (with project number 452019308) of the research programme Rubicon which is (partly) financed by the Dutch Research Council (NWO). This research is partly funded by the Amsterdam Cardiovascular Sciences (personal grant F.V.Y.T).