In many modern machine learning applications, changes in covariate distributions and difficulty in acquiring outcome information have posed challenges to robust model training and evaluation. Numerous transfer learning methods have been developed to robustly adapt the model itself to some unlabeled target populations using existing labeled data in a source population. However, there is a paucity of literature on transferring performance metrics, especially receiver operating characteristic (ROC) parameters, of a trained model. In this paper, we aim to evaluate the performance of a trained binary classifier on unlabeled target population based on ROC analysis. We proposed Semisupervised Transfer lEarning of Accuracy Measures (STEAM), an efficient three-step estimation procedure that employs (1) double-index modeling to construct calibrated density ratio weights and (2) robust imputation to leverage the large amount of unlabeled data to improve estimation efficiency. We establish the consistency and asymptotic normality of the proposed estimator under the correct specification of either the density ratio model or the outcome model. We also correct for potential overfitting bias in the estimators in finite samples with cross-validation. We compare our proposed estimators to existing methods and show reductions in bias and gains in efficiency through simulations. We illustrate the practical utility of the proposed method on evaluating prediction performance of a phenotyping model for rheumatoid arthritis (RA) on a temporally evolving EHR cohort.

Predicting sets of outcomes-instead of unique outcomes-is a promising solution to uncertainty quantification in statistical learning. Despite a rich literature on constructing prediction sets with statistical guarantees, adapting to unknown covariate shift-a prevalent issue in practice-poses a serious unsolved challenge. In this article, we show that prediction sets with finite-sample coverage guarantee are uninformative and propose a novel flexible distribution-free method, PredSet-1Step, to efficiently construct prediction sets with an asymptotic coverage guarantee under unknown covariate shift. We formally show that our method is asymptotically probably approximately correct, having well-calibrated coverage error with high confidence for large samples. We illustrate that it achieves nominal coverage in a number of experiments and a data set concerning HIV risk prediction in a South African cohort study. Our theory hinges on a new bound for the convergence rate of the coverage of Wald confidence intervals based on general asymptotically linear estimators.

BACKGROUND: Concept drift and covariate shift lead to a degradation of machine learning (ML) models. The objective of our study was to characterize sudden data drift as caused by the COVID pandemic. Furthermore, we investigated the suitability of certain methods in model training to prevent model degradation caused by data drift.
METHODS: We trained different ML models with the H2O AutoML method on a dataset comprising 102,666 cases of surgical patients collected in the years 2014-2019 to predict postoperative mortality using preoperatively available data. Models applied were Generalized Linear Model with regularization, Default Random Forest, Gradient Boosting Machine, eXtreme Gradient Boosting, Deep Learning and Stacked Ensembles comprising all base models. Further, we modified the original models by applying three different methods when training on the original pre-pandemic dataset: (Rahmani K, et al, Int J Med Inform 173:104930, 2023) we weighted older data weaker, (Morger A, et al, Sci Rep 12:7244, 2022) used only the most recent data for model training and (Dilmegani C, 2023) performed a z-transformation of the numerical input parameters. Afterwards, we tested model performance on a pre-pandemic and an in-pandemic data set not used in the training process, and analysed common features.
RESULTS: The models produced showed excellent areas under receiver-operating characteristic and acceptable precision-recall curves when tested on a dataset from January-March 2020, but significant degradation when tested on a dataset collected in the first wave of the COVID pandemic from April-May 2020. When comparing the probability distributions of the input parameters, significant differences between pre-pandemic and in-pandemic data were found. The endpoint of our models, in-hospital mortality after surgery, did not differ significantly between pre- and in-pandemic data and was about 1% in each case. However, the models varied considerably in the composition of their input parameters. None of our applied modifications prevented a loss of performance, although very different models emerged from it, using a large variety of parameters.
CONCLUSIONS: Our results show that none of our tested easy-to-implement measures in model training can prevent deterioration in the case of sudden external events. Therefore, we conclude that, in the presence of concept drift and covariate shift, close monitoring and critical review of model predictions are necessary.

The excellent performance of deep neural networks on image classification tasks depends on a large-scale high-quality dataset. However, the datasets collected from the real world are typically biased in their distribution, which will lead to a sharp decline in model performance, mainly because an imbalanced distribution results in the prior shift and covariate shift. Recent studies have typically used a two-stage learning method consisting of two rebalancing strategies to solve these problems, but the combination of partial rebalancing strategies will damage the representational ability of the networks. In addition, the two-stage learning method is of little help in addressing the problem of covariate shift. To solve the above two issues, we first propose a sample logit-aware reweighting method called (SLA), which can not only repair the weights of majority class hard samples and minority class samples but will also integrate with logit adjustment to form a stable two-stage learning strategy. Second, to solve the covariate shift problem, inspired by ensemble learning, we propose a multi-domain expert specialization model, which can achieve a more comprehensive decision by averaging expert classification results from multiple different domains. Finally, we combine SLA and logit adjustment into a two-stage learning method and apply our model to the CIFAR-LT and ImageNet-LT datasets. Compared with the most advanced methods, our experimental results show excellent performance.

Individualized treatment effect lies at the heart of precision medicine. Interpretable individualized treatment rules (ITRs) are desirable for clinicians or policymakers due to their intuitive appeal and transparency. The gold-standard approach to estimating the ITRs is randomized experiments, where subjects are randomized to different treatment groups and the confounding bias is minimized to the extent possible. However, experimental studies are limited in external validity because of their selection restrictions, and therefore the underlying study population is not representative of the target real-world population. Conventional learning methods of optimal interpretable ITRs for a target population based only on experimental data are biased. On the other hand, real-world data (RWD) are becoming popular and provide a representative sample of the target population. To learn the generalizable optimal interpretable ITRs, we propose an integrative transfer learning method based on weighting schemes to calibrate the covariate distribution of the experiment to that of the RWD. Theoretically, we establish the risk consistency for the proposed ITR estimator. Empirically, we evaluate the finite-sample performance of the transfer learner through simulations and apply it to a real data application of a job training program.

Machine-learning models are susceptible to external influences which can result in performance deterioration. The aim of our study was to elucidate the impact of a sudden shift in covariates, like the one caused by the Covid-19 pandemic, on model performance.
After ethical approval and registration in Clinical Trials (NCT04092933, initial release 17/09/2019), we developed different models for the prediction of perioperative mortality based on preoperative data: one for the pre-pandemic data period until March 2020, one including data before the pandemic and from the first wave until May 2020, and one that covers the complete period before and during the pandemic until October 2021. We applied XGBoost as well as a Deep Learning neural network (DL). Performance metrics of each model during the different pandemic phases were determined, and XGBoost models were analysed for changes in feature importance.
XGBoost and DL provided similar performance on the pre-pandemic data with respect to area under receiver operating characteristic (AUROC, 0.951 vs. 0.942) and area under precision-recall curve (AUPR, 0.144 vs. 0.187). Validation in patient cohorts of the different pandemic waves showed high fluctuations in performance from both AUROC and AUPR for DL, whereas the XGBoost models seemed more stable. Change in variable frequencies with onset of the pandemic were visible in age, ASA score, and the higher proportion of emergency operations, among others. Age consistently showed the highest information gain. Models based on pre-pandemic data performed worse during the first pandemic wave (AUROC 0.914 for XGBoost and DL) whereas models augmented with data from the first wave lacked performance after the first wave (AUROC 0.907 for XGBoost and 0.747 for DL). The deterioration was also visible in AUPR, which worsened by over 50% in both XGBoost and DL in the first phase after re-training.
A sudden shift in data impacts model performance. Re-training the model with updated data may cause degradation in predictive accuracy if the changes are only transient. Too early re-training should therefore be avoided, and close model surveillance is necessary.

Causally interpretable meta-analysis combines information from a collection of randomized controlled trials to estimate treatment effects in a target population in which experimentation may not be possible but from which covariate information can be obtained. In such analyses, a key practical challenge is the presence of systematically missing data when some trials have collected data on one or more baseline covariates, but other trials have not, such that the covariate information is missing for all participants in the latter. In this article, we provide identification results for potential (counterfactual) outcome means and average treatment effects in the target population when covariate data are systematically missing from some of the trials in the meta-analysis. We propose three estimators for the average treatment effect in the target population, examine their asymptotic properties, and show that they have good finite-sample performance in simulation studies. We use the estimators to analyze data from two large lung cancer screening trials and target population data from the National Health and Nutrition Examination Survey (NHANES). To accommodate the complex survey design of the NHANES, we modify the methods to incorporate survey sampling weights and allow for clustering.

We propose methods for estimating the area under the receiver operating characteristic (ROC) curve (AUC) of a prediction model in a target population that differs from the source population that provided the data used for original model development. If covariates that are associated with model performance, as measured by the AUC, have a different distribution in the source and target populations, then AUC estimators that only use data from the source population will not reflect model performance in the target population. Here, we provide identification results for the AUC in the target population when outcome and covariate data are available from the sample of the source population, but only covariate data are available from the sample of the target population. In this setting, we propose three estimators for the AUC in the target population and show that they are consistent and asymptotically normal. We evaluate the finite-sample performance of the estimators using simulations and use them to estimate the AUC in a nationally representative target population from the National Health and Nutrition Examination Survey for a lung cancer risk prediction model developed using source population data from the National Lung Screening Trial.

We propose a direct domain adaptation (DDA) approach to enrich the training of supervised neural networks on synthetic data by features from real-world data. The process involves a series of linear operations on the input features to the NN model, whether they are from the source or target distributions, as follows: (1) A cross-correlation of the input data (i.e., images) with a randomly picked sample pixel (or pixels) of all images from the input or the mean of all randomly picked sample pixel (or pixels) of all input images. (2) The convolution of the resulting data with the mean of the autocorrelated input images from the other domain. In the training stage, as expected, the input images are from the source distribution, and the mean of auto-correlated images are evaluated from the target distribution. In the inference/application stage, the input images are from the target distribution, and the mean of auto-correlated images are evaluated from the source distribution. The proposed method only manipulates the data from the source and target domains and does not explicitly interfere with the training workflow and network architecture. An application that includes training a convolutional neural network on the MNIST dataset and testing the network on the MNIST-M dataset achieves a 70% accuracy on the test data. A principal component analysis (PCA), as well as t-SNE, shows that the input features from the source and target domains, after the proposed direct transformations, share similar properties along the principal components as compared to the original MNIST and MNIST-M input features.

We considered methods for transporting a prediction model for use in a new target population, both when outcome and covariate data for model development are available from a source population that has a different covariate distribution compared with the target population and when covariate data (but not outcome data) are available from the target population. We discuss how to tailor the prediction model to account for differences in the data distribution between the source population and the target population. We also discuss how to assess the model\'s performance (e.g., by estimating the mean squared prediction error) in the target population. We provide identifiability results for measures of model performance in the target population for a potentially misspecified prediction model under a sampling design where the source and the target population samples are obtained separately. We introduce the concept of prediction error modifiers that can be used to reason about tailoring measures of model performance to the target population. We illustrate the methods in simulated data and apply them to transport a prediction model for lung cancer diagnosis from the National Lung Screening Trial to the nationally representative target population of trial-eligible individuals in the National Health and Nutrition Examination Survey.