Introduction
Machine learning models in healthcare often struggle with the "research-to-clinical" gap, where performance degrades significantly when moving from a controlled training environment to real-world clinical settings. This phenomenon, known as domain shift or dataset bias, arises from variations in patient demographics, laboratory equipment, and local disease prevalence. While standard Domain Adaptation (DA) techniques exist, they typically require prior knowledge of or data from the specific target domain. In urgent scenarios like the COVID-19 pandemic, "target-free" models—those capable of immediate deployment in unseen environments—are essential. This paper presents "Cross-Adaptation," a methodology designed to enhance model generalization without requiring target-site data.
Methodology: The Cross-Adaptation Algorithm
The proposed Cross-Adaptation (CA) method is a data-level preprocessing step that iteratively aligns multiple source datasets to create a more robust training set. Unlike simple data pooling, CA transforms each individual source dataset to match the distribution of the collective "rest of the data."
The process follows four primary steps:
- Iterative Alignment: For a collection of $n$ datasets ($D_1, D_2, \dots, D_n$), one dataset is designated as the Source ($D_s$), while the union of the remaining $n-1$ datasets serves as the Target ($D_t$).
- Transformation: A domain adaptation function $g(D_s, D_t)$ is applied to transform $D_s$ so its distribution aligns with $D_t$. The authors evaluated several DA methods, including Kernel Mean Matching (KMM), Kullback-Leibler Importance Estimation Procedure (KLIEP), and Transfer AdaBoost.
- Concatenation: This procedure is repeated for every dataset in the collection. The final training set is the concatenation of all transformed versions.
- Model Training: Because CA operates at the data level, it is classifier-agnostic, allowing the resulting data to be used with any machine learning architecture.
Experimental Setup
The effectiveness of Cross-Adaptation was evaluated using SARS-CoV-2 detection based on Complete Blood Count (CBC) tests. CBC tests are routine and inexpensive, making them ideal for rapid screening. The study utilized 9 datasets comprising 4,870 tests from hospitals across Brazil, Italy, Poland, Ethiopia, and Spain. Each test included 10 standard parameters (e.g., white blood cell count, hemoglobin, leukocyte percentages) and sex, with RT-PCR results serving as the ground truth.
Results
The researchers compared the performance of five machine learning models—K-Nearest Neighbors (KNN), Decision Trees, Random Forests, XGBoost, and Multi-Layer Perceptrons (MLP)—trained on either raw untransformed source data or cross-adapted data.
As shown in the results, Cross-Adaptation consistently outperformed the baseline across all architectures. On average, the method yielded a 10-percentage-point increase in F1 scores. Notably, models like KNN and XGBoost saw improvements from 55% and 62% to 73%, respectively. These results suggest that by smoothing out site-specific biases during the training phase, the models learn features that are more invariant and thus more generalizable to new, unseen clinical environments.
Conclusion and Future Work
Cross-Adaptation offers a promising, architecture-independent solution to dataset bias in medical AI. By leveraging the diversity of existing source datasets to simulate a generalized target distribution, it enables the creation of robust models ready for immediate deployment. Future research will focus on the computational scalability of the iterative process as the number of source domains increases and the application of this logic to more complex data modalities, such as medical imaging.
Summary updated