2402.01987

# Online Transfer Learning for RSV Case Detection **Authors**: Yiming Sun, Yuhe Gao, Runxue Bao, Gregory F. Cooper{@IEEEauthorhalign}Jessi Espino, Harry Hochheiser, Marian G. Michaels{@IEEEauthorhalign}John M. Aronis, Chenxi Song, Ye Ye†† † > Department of Electrical andComputer EngineeringUniversity of PittsburghPittsburgh, USAyis108@pitt.edu > Department ofBiomedical InformaticsUniversity of PittsburghPittsburgh, USAyug51@pitt.edu > Department of Electrical andComputer EngineeringUniversity of PittsburghPittsburgh, USArunxue.bao@pitt.edu > Department ofBiomedical InformaticsUniversity of PittsburghPittsburgh, USAgfc@pitt.eduDepartment of Biomedical InformaticsUniversity of PittsburghPittsburgh, USAjue@pitt.edu > Department of Biomedical InformaticsUniversity of PittsburghPittsburgh, USAharryh@pitt.edu > Department of PediatricsUniversity of PittsburghPittsburgh, USAmichmg@pitt.eduDepartment of Biomedical InformaticsUniversity of PittsburghPittsburgh, USAjma18@pitt.edu > Department of Biomedical InformaticsUniversity of PittsburghPittsburgh, USAchs342@pitt.edu > ††\dagger†Corresponding authorDepartment of Biomedical InformaticsUniversity of PittsburghPittsburgh, USAyey5@pitt.edu Abstract Transfer learning has become a pivotal technique in machine learning and has proven to be effective in various real-world applications. However, utilizing this technique for classification tasks with sequential data often faces challenges, primarily attributed to the scarcity of class labels. To address this challenge, we introduce Multi-Source Adaptive Weighting (MSAW), an online multi-source transfer learning method. MSAW integrates a dynamic weighting mechanism into an ensemble framework, enabling automatic adjustment of weights based on the relevance and contribution of each source (representing historical knowledge) and target model (learning from newly acquired data). We demonstrate the effectiveness of MSAW by applying it to detect Respiratory Syncytial Virus cases within Emergency Department visits, utilizing multiple years of electronic health records from the University of Pittsburgh Medical Center. Our method demonstrates performance improvements over many baselines, including refining pre-trained models with online learning as well as three static weighting approaches, showing MSAW’s capacity to integrate historical knowledge with progressively accumulated new data. This study indicates the potential of online transfer learning in healthcare, particularly for developing machine learning models that dynamically adapt to evolving situations where new data is incrementally accumulated. Index Terms: online transfer learning, ensemble method, dynamic weighting mechanism, electronic health record, Respiratory Syncytial Virus case detection I Introduction Machine learning has made substantial advancements in recent decades, with its applications spanning a wide range of fields such as computer vision (CV), natural language processing (NLP), speech recognition, and autonomous driving. Despite these achievements, machine learning in biomedicine faces significant challenges, particularly in data collection. The acquisition of labeled data can be very costly or even infeasible due to factors like ethical considerations, patient privacy, and the scarcity of certain diseases. These challenges have led researchers to increasingly rely on utilizing data from related domains that have a more abundant supply of data. To better leverage knowledge from these related yet distinct domains, transfer learning has emerged as a promising approach to enhance model performance. Fundamental to transfer learning are the concepts of the ‘source domain’ and ‘target domain’. The source domain, which is usually rich in data, serves as the foundation for knowledge extraction. It is characterized by its abundance of well-labeled datasets, which aid models in understanding and encoding complex patterns. In contrast, the target domain, where this pre-learned knowledge is used, often features a more limited dataset. Transfer learning bridges this gap by adapting models trained on the source domain to function effectively in the target domain, thereby helping address the challenges associated with data scarcity. Traditional transfer learning approaches [1, 2, 3] typically depend on having complete datasets from both the source and target domains at the start of the training process. However, this is often not feasible, especially in dynamic environments where target domain data are collected progressively. To tackle this issue, the concept of online transfer learning [4, 5, 6] has been introduced. This method is designed for scenarios where target domain data are acquired incrementally over time, while source domain data are already accessible before training begins. In such cases, transferring knowledge from the source domain becomes crucial, particularly because the limited initial data in the target domain may be insufficient for effective learning. The extensive and diverse information available from the source domains can significantly compensate for this shortfall, providing an initial knowledge base that the model can build upon as more target domain data becomes available. Therefore, the efficiency and effectiveness of learning in the target domain are greatly enhanced by the transferred knowledge from the source domains. Online transfer learning entails leveraging knowledge from a static source domain and applying it to an ongoing, evolving target task. The model is incrementally updated as new target data are received. A crucial element of this process is the utilization of a loss function to fine-tune the model in accordance with the latest incoming data, thus ensuring its continued relevance and efficiency. This adaptability renders online transfer learning particularly suitable for real-time applications, effectively handling issues such as concept drift [7] and providing a more versatile approach to knowledge transfer. In traditional transfer learning, the distinction between various domains is clear and easily identifiable. This clarity allows for straightforward categorization based on common characteristics, such as grouping images by distinguishing features like those of cats versus dogs. However, online transfer learning presents a nuanced challenge as it focuses on the evolution of a single task over time, making the determination of where to apply temporal segmentation less clear. This difficulty arises because the task’s concept remains unchanged, while the data representing the task may significantly change due to temporal factors. This evolution can alter the relevance of past information to current or future tasks, making it difficult to understand how these varying time periods connect and influence one another. Unlike static categorizations, where similarities and differences are often clear and measurable, the dynamic nature of temporal domains means that the relevance of past data to the present task can be fluid and elusive. For instance, in healthcare, temporal segmentation might involve dividing patient data into annual periods to capture seasonal disease trends. With multiple source domains, integrating them optimally into an ensemble model is challenging. A critical question arises: how do we determine which source domains are most relevant to the current period (or target domain) and which are less relevant or unrelated? Our proposed method employs a sequential data acquisition strategy to tackle this challenge. Consider a healthcare scenario where the task is predicting patient outcomes in a new medical facility. To address initial data scarcity in this facility, our approach significantly relies on pre-trained models derived from source domain data, such as historical patient outcomes from established facilities. These pre-trained source models form the initialization of our ensemble model. This ensemble is adept at autonomously evaluating and determining the relevance of each source model’s learned patterns to the current patient data in the new facility (target model). For instance, a source model trained on a large urban hospital might be initially prioritized for its broad spectrum of cases. As more patient data accumulates in the new facility, the ensemble dynamically adjusts its reliance on this urban model, potentially shifting focus to another source model that aligns more closely with emerging local trends. This strategic utilization of pre-trained source models ensure that the most effective ones are prioritized during the critical early stages of training in the target domain. As the target domain accumulates more data, the ensemble model modifies its approach. Initially dependent on pre-trained source models, it gradually begins to prioritize the continually updated target model. This change is twofold: a shift in reliance from source to target models and a nuanced adjustment in the weight each model holds in decision-making. This transition enables the model to evolve alongside the growing target domain dataset. The distinct roles of shifting focus and adjusting weights demonstrate the adaptability of our online transfer learning approach, particularly in environments where data are collected over time. Our framework makes two main contributions. First, we introduce an online multi-source transfer learning architecture for machine learning using data over multiple years. Second, we explore the performance of different ways of generating weights for the ensemble model to improve its accuracy. We evaluate the effectiveness of our approach using real-world Respiratory Syncytial Virus (RSV) patient encounter data that were collected over multiple years from Emergency Departments (EDs) at University of Pittsburgh Medical Center (UPMC). Our primary target-domain goal is to utilize patients’ signs and symptoms to accurately classify them as either having RSV or not, a critical step towards improving infectious disease case detection using EHRs. RSV is a major cause of respiratory infections that can lead to hospitalizations [8]. The burden of RSV continues to grow, especially in young children aged less than 5 years old [9, 10]. RSV also has been identified as an important disease causing hospitalizations of the elderly and high-risk adults [11]. The disease typically exhibits a seasonal pattern in case numbers [12]. In this study, we segment the RSV case data into distinct domains based on a defined annual cycle, with each domain spanning from June of the previous year to May of the current year. This segmentation strategy is crucial to our framework, aiming to improve RSV case detection accuracy by using data from multiple preceding seasons as source inputs. For instance, we use patient encounter data from June 2011 to May 2018 to develop an ensemble model for classifying encounters from June 2018 to May 2019. The experimental results indicate that our approach improves classification performance, particularly related to public health planning and resource allocation for anticipated RSV seasons. II Related Work Traditional transfer learning methodologies typically assume that the source and target domains are predefined, aiming to transfer knowledge from the source to the target domain. Many comprehensive surveys [13, 14, 15, 16, 17] have detailed and categorized a broad range of transfer learning approaches, evaluating factors such as algorithm types and application scenarios. A notable recent development is the growing body of literature reviews focusing on the integration of transfer learning with other advanced machine learning techniques. These include combinations with deep learning [18, 19], reinforcement learning [20, 21], and federated learning [22, 23]. Additionally, some reviews have specifically targeted the application of transfer learning in key areas such as CV [24, 25], NLP [26, 27, 28], healthcare [29, 30, 31, 32, 33], and wireless communication [34, 35], providing insights into the practical implications and utility of transfer learning in these specific fields. Online transfer learning has emerged to address situations where data are received sequentially and are not available at the start of training. This approach is particularly relevant in dynamic environments where real-time data processing is crucial. Some popular online transfer learning methods [36, 37, 38, 39] have demonstrated excellent performance in many real-world scenarios. Online transfer learning methods can be categorized into data-based methods and model-based methods. Data-based transfer learning methods involve the transfer of original data to a target domain, allowing the target models to be trained with these augmented data, thereby enriching the available data within the target domain. Online Heterogeneous Knowledge Transition [36] trains the target model using data that co-occur in source and target domains. The co-occurrence data can be a set of text-image pairs, for example, the image includes a red car and the text is ‘red; wheel; transportation’. Conversely, model-based transfer learning methods focus on firstly building models within the source domain, then adapting both the model structure and parameters for the target domain. Here, ‘model structure’ refers to the architecture or configuration of the model (such as layers in a neural network), while ‘model parameters’ are the specific coefficients that are learned from the source data. The target model leverages insights from the source domain to potentially improve performance. Ensemble models, which combine the outputs or features of multiple models, are commonly employed in model-based methods to further improve accuracy and robustness [37, 38]. Online Heterogeneous Transfer by Hedge Ensemble [39] handles the online heterogeneous transfer problem by updating weights of offline and online classifiers based on the Hedge weighting strategy. For our online transfer learning algorithm, we used RSV detection as a representative example task. The use of machine learning for RSV detection has been growing. Tso et al. [40] developed an extreme gradient boost decision tree model to detect RSV cases in both emergency department and inpatient settings, using clinical measurements from EHRs, including respiratory rate and peripheral oxygen saturation. In another study, Cord et al. [41] assessed different methodologies such as Logistic Regression, Gaussian Naive Bayes, and Random Forest, ultimately choosing the Light Gradient Boosting Machine algorithm for their RSV detection models. III Methodology In this section, we begin by outlining the problem formulation for the online multiple source-domain scenario. Following this, we delve into the base model, setting the stage for the introduction of our proposed method. III-A Problem Formulation We address the challenge of binary classification in environments where data are dynamic, arriving sequentially and often associated with specific time frames such as epidemiological seasons. Our focus is on creating a method to analyze information collected across seasons, using Bayesian networks as the basis of our classifier training. In practical scenarios, data from each season can be viewed as a separate domain. This temporal aspect presents a distinct challenge, especially at the beginning of a new season when data are scarce. To tackle this problem, we suggest a transfer learning strategy where the current season is the target domain, and data from several past seasons are used as source domains. In transfer learning, research [42] has indicated that simply combining multiple source domains into one is inferior to employing a multi-source strategy. Thus, our method uses data from multiple source domains, represented as $D^{S}=\{D_{S}^{1},D_{S}^{2},...,D_{S}^{n}\}$ , where each $D_{S}^{i}$ is data from a specific season, together with a single target domain $D_{T}$ . We leverage historical data from past seasons, assuming that patterns and knowledge can effectively inform and improve the classification model for the current season. III-B Base Model Our experiment study begins with the implementation of the Naive Bayes classifiers. The effectiveness of Naive Bayes stems from its simplicity, which facilitates tracking model changes across seasons over time, and its computational efficiency, particularly when the features are relatively independent given the class label. An additional benefit of the Naive Bayes model is its statistical efficiency, which allows one to construct a model from a small number of instances, making it particularly well-suited for online learning with only target data. III-C Online Multi-source Ensemble Model In our framework, for each source domain $D_{S}^{i}$ , we train an offline base model $f_{S}^{i}$ as a classifier on this domain. Similarly, a specific online training classifier $f_{T}$ is designated for the target domain. The architecture of our proposed model is depicted in Fig. 1. Central to our approach is an ensemble model, formulated as: $$ f=w_{1}f_{S}^{1}+\ldots+w_{n}f_{S}^{n}+w_{T}f_{T} \tag{1} $$ Here, $w_{i}$ represents the weight parameter for the $i$ -th source domain classifier, and $w_{T}$ denotes the weight parameter for the target domain classifier. Each source domain classifier $f_{S}^{i}$ is pre-trained on its respective domain data, while the target domain classifier $f_{T}$ is dynamically trained and updated with new instances from the target domain as they become available. <details> <summary>2402.01987v2/x1.png Details</summary>  ### Visual Description # Technical Document Extraction: Flowchart Analysis ## Diagram Overview The image depicts a **multi-stage classification system architecture** with the following key components: ### 1. Source Classifiers (D_S^1 to D_S^n) - **Labels**: - `D_S^1`, `D_S^2`, ..., `D_S^n` (source classifiers) - `D_T` (target classifier) - **Structure**: - Each classifier is represented as an oval node - Connected via downward arrows to corresponding function blocks ### 2. Function Blocks (f_S^1 to f_S^n, f_T) - **Labels**: - `f_S^1`, `f_S^2`, ..., `f_S^n` (source function blocks) - `f_T` (target function block) - **Weights**: - `w_1`, `w_2`, ..., `w_n` (weights for source functions) - `w_T` (weight for target function) - **Visual Representation**: - Rectangular blocks with labels - Positioned directly below corresponding classifiers ### 3. Aggregation Mechanism - **Symbol**: Σ (sigma) - **Function**: - Represents summation/aggregation of weighted function outputs - Connects all function blocks to the final classifier ### 4. Final Classifier - **Label**: "Final Classifier" - **Position**: - Located at the bottom center - Receives aggregated input from all preceding components ## Flowchart Logic 1. **Input Flow**: - Each source classifier (`D_S^i`) processes input through its function block (`f_S^i`) - Target classifier (`D_T`) processes input through `f_T` 2. **Weighted Contribution**: - Each function output is scaled by its respective weight (`w_i` or `w_T`) 3. **Aggregation**: - Weighted outputs are summed via the Σ operator 4. **Output**: - Final classifier synthesizes aggregated information for final decision ## Technical Notes - **Notation**: - Superscripts (¹, ², ..., ⁿ) indicate sequential stages - Subscripts (S, T) denote source vs. target components - **Architecture Type**: - Ensemble learning framework - Combines multiple classifiers with weighted contributions - **Missing Elements**: - No explicit data values or numerical trends present - No temporal or spatial dimensions shown ## Diagram Components Summary | Component Type | Count | Labels | |----------------------|-------|---------------------------------| | Source Classifiers | n | D_S^1, D_S^2, ..., D_S^n | | Target Classifier | 1 | D_T | | Function Blocks | n+1 | f_S^1, f_S^2, ..., f_S^n, f_T | | Weights | n+1 | w_1, w_2, ..., w_n, w_T | | Aggregation Symbol | 1 | Σ | | Final Output | 1 | Final Classifier | ## Spatial Relationships - **Vertical Alignment**: - Classifiers → Function Blocks → Final Classifier - **Horizontal Grouping**: - Source components (D_S^i, f_S^i, w_i) grouped left - Target component (D_T, f_T, w_T) isolated right - **Aggregation Position**: - Σ centered below all function blocks </details> Figure 1: Architecture of the proposed ensemble model. Our ensemble method capitalizes on the strengths of individual classifiers trained on distinct historical data patterns from each source domain, integrating these with the versatile classifier of the target domain. This integration aims to enhance the model’s overall classification capabilities and accuracy, which is especially beneficial when dealing with limited initial data in the target domain. For weight allocation in our ensemble model, adaptive weights can provide a dynamic adjustment mechanism. Input: Pre-trained source models $f_{S}^{1}(·),...,f_{S}^{n}(·)$ , Target model $f_{T}(·)$ , Target Dataset $X=\{X_{j}\}$ for $j=1,2,...$ , Weight penalty $\alpha∈\left(1,∞\right)$ , Target weight factor $\beta∈\left(0,1\right)$ . Initialization: Initialize source model weights $w_{S}^{i}=1.0$ for $i=1,2,...,n$ target model weight $w_{T}=1.0$ , and current data volume in target domain $j$ . while receive new data point $X_{j}$ do Increase target data volume $j← j+1$ Adjust target weight: $w_{T}← w_{T}×\beta× j$ Normalize weights: $\displaystyle w_{S}^{i}←\frac{w_{S}^{i}}{\sum_{i=1}^{n}w_{S}^{i}+w_{T% }},$ $\displaystyle w_{T}←\frac{w_{T}}{\sum_{i=1}^{n}w_{S}^{i}+w_{T}}$ Get combined probability for final classification: $prob=\sum_{i=1}^{n}w_{S}^{i}× f_{S}^{i}(X_{j})+w_{T}× f_{T}(X_{j})$ for $i=1$ to $n$ do if $f_{S}^{i}(X_{j})$ classifies incorrectly then Adjust weight: $w_{S}^{i}← w_{S}^{i}×\frac{\sqrt{j}}{\sqrt{j}+\alpha}$ end if end for if $f_{T}(X_{j})$ classifies incorrectly then Adjust weight: $w_{T}← w_{T}×\frac{\sqrt{j}}{\sqrt{j}+\alpha}$ end if Update $f_{T}$ with $X_{j}$ Return $prob$ as the classifying probability for $X_{j}$ end while Algorithm 1 Multi-Source Adaptive Weighting (MSAW) Our proposed Multi-Source Adaptive Weighting (MSAW) method, detailed in Algorithm 1, involves updating all weights, covering those of the target domain classifier and source domain classifiers. As new training instances are processed, the algorithm assesses the accuracy of each classifier, adjusting weight based on their classifying performance. Weights of accurate classifiers are unchanged while those of inaccurate classifiers are penalized, diminishing their impact on future classifications. MSAW also dynamically adjusts the weights of classifiers based on changes in data volume. As the amount of target domain data increases, the target classifier’s weight is proportionally enhanced. This increase is quantified by multiplying the weight with a factor that signifies the quantity of target data, a process facilitated by the variable $j$ . The variable $j$ serves as a sequential index for each new target data point, effectively tracking the expansion of the target dataset. As more target instances are received (i.e., as $j$ increases), the target classifier’s weight is correspondingly adjusted upwards. This adjustment ensures that the weight of the target classifier is attuned to the scale of the available data, enabling the algorithm to progressively prioritize more recent and relevant insights. Simultaneously, the source domain classifiers’ weights are adjusted primarily based on their classification performance. This dual mechanism—adapting weights for both classification accuracy and data volume—enables a more dynamic and responsive approach in the MSAW algorithm. The target classifier’s weight becomes increasingly significant as more target data become available, ensuring that the model’s classifications are continually refined and aligned with the latest trends and patterns in the target domain. We hypothesize that the adaptive nature of the learnable weight assignment strategy in the proposed algorithm will outperform static weight assignment strategies. This hypothesis rests on the idea that dynamically adjusting weights enables the ensemble model to evolve in sync with incoming target data, potentially leading to more precise and adaptable classifications. The three static weighting assignment strategies for comparison are: (1) The ‘Equal Weights’ method which assigns each source model a constant and equal weight. (2) The ‘Volume-based’ method assigns weights to source models based on the size of their training datasets. (3) The ‘Time-based’ method combines individual models based on their temporal proximity to the testing season (target domain), where ‘time distance’ is defined as the number of seasons between a model’s training data and the testing period. IV Dataset The research data were sourced from the University of Pittsburgh’s Neptune Research Data Warehouse [43], and the study protocol received approval from the University of Pittsburgh Institutional Review Board (Protocol STUDY20030193). The dataset consists of 2,263,360 encounters from five UPMC EDs (Presbyterian University Hospital, Children’s Hospital, McKeesport Hospital, Mercy Hospital, and Shadyside Hospital) in Allegheny County over nine years from January 2011 to May 2020. In regions with temperate climates, RSV infections predominantly manifest during the colder winter months. The onset of these infections is usually observed around October and November in the Northern Hemisphere. Typically, the highest incidence of RSV cases, or the peak period, occurs in January and February. As the season progresses, the frequency of RSV infections starts to decline, with the final cases of the season generally being recorded between March and May [12]. Given the seasonal nature of RSV outbreaks, we defined one season as spanning from June of one year to May of the following year, rather than using a calendar year division for data segmentation. Table I details the sizes of encounters for each season. An encounter was classified as related to RSV if it included one of any of the following International Classification of Diseases (ICD) codes at discharge; encounters that did not include any of these ICD codes were categorized as non-RSV. - 79.6 Respiratory syncytial virus - 466.11 Acute bronchiolitis due to respiratory syncytial virus - 480.1 Pneumonia due to respiratory syncytial virus - B97.4 Respiratory syncytial virus as the cause of diseases classified elsewhere - J12.1 Respiratory syncytial virus pneumonia - J20.5 Acute bronchitis due to respiratory syncytial virus - J21.0 Acute bronchiolitis due to respiratory syncytial virus TABLE I: Sizes of datasets | Season | Number of RSV Labeled Cases | Number of Non-RSV Labeled Cases | | --- | --- | --- | | 2011-2012 | 251 | 259,596 | | 2012-2013 | 238 | 258,024 | | 2013-2014 | 132 | 254,076 | | 2014-2015 | 179 | 250,476 | | 2015-2016 | 200 | 252,432 | | 2016-2017 | 381 | 257,172 | | 2017-2018 | 408 | 243,708 | | 2018-2019 | 507 | 240,312 | | 2019-2020 | 888 | 244,380 | Our dataset contains 7,614 features, including demographic information, lab test results, and NLP-extracted clinical findings, detailed in Table II. The clinical findings were obtained by using the MetaMap parser [44], which applies pattern-matching and deduction rules. The clinical notes for each encounter serve as input for the parser, which extracts a set of clinical findings and their values. The values for each finding are either present (P), negated (N), or missing (M), with missing denoting that clinicians did not mention the finding in the note, as per the parser’s output. The recommendation to designate missing values in NLP-retrieved findings from EHRs as a distinct category was made by [45]. In cases where an encounter had multiple clinical notes, we used the union of all clinical findings across all the notes. Contradictions across notes were resolved in favor of present findings, and this also applied to multiple occurrences of a finding in a single note. TABLE II: Features in research datasets | Features | Distinct Values (Descriptions) | | --- | --- | | age | less5 ( $<$ 5), ge5less18 (5-17), ge18less65 (8-64), ge65 (65+), M (value is missing) | | gender | female, male, unknown, M (value is missing) | | ethnicity | Hispanic or Latino, not Hispanic or Latino, not specified, M (value is missing) | | race | black, white, other, not specified, M (value is missing) | | 7541 UMLS CUIs extracted from free-text reports by MetaMap | P (present), N (negated), and M (value is missing). One CUI example is C0859927 (Increased work of breathing). | | 69 laboratory results | H (high), L (low), N (normal), M (value is missing). One lab example is oxygen saturation. | V Experiments & Results V-A Experiment Settings We employed an online multi-source transfer learning ensemble framework to enhance model’s adaptability in a specific target domain. This framework incorporated datasets from previous seasons, each treated as a distinct source domain. We evaluated models’ performance using the Area Under the Receiver Operating Characteristic Curve (AUROC), an important measure for gauging a model’s discriminative performance when classifying positive and negative cases. In the specified algorithm, two hyperparameters, $\alpha$ and $\beta$ , play crucial roles in the dynamic adjustment of model weights during the learning process. The setting of $\alpha$ and $\beta$ is consistent across all domains. The hyperparameter $\alpha$ is set to $\log_{2}10$ = 3.32, a value selected after experimenting with several hyperparameter settings. This value for $\alpha$ was chosen to balance the weight penalty applied to a classifier’s weights upon making an incorrect classification. This particular value for $\alpha$ strikes a balance between being too lenient and too punitive, allowing for a gradual decrease in a classifier’s influence while still retaining its contributions where they prove useful. The hyperparameter $\beta$ , set at $1/200000$ , plays a critical role in the volume-adaptive component of our algorithm. This parameter is instrumental in fine-tuning the adjustment of the target weight $w_{T}$ , aligning it with the growing volume of incoming data. Here, the factor $j$ serves as an index representing the sequence of data points being processed. The specific choice of $1/200000$ for $\beta$ strikes a delicate balance. When the number of instances has not yet reached 200,000, this configuration ensures that the target weights are subtly suppressed, preventing premature dominance in the model’s decision-making process. However, once the count surpasses this threshold, the target weights are progressively amplified, reflecting the increased reliability and relevance of the more substantial data corpus. This ensures that the target model’s influence is appropriately scaled, growing more pronounced as the volume of target data becomes more significant, thereby sensitively responding to the increase in data volume while simultaneously tempering the rate of weight adjustment. By adopting this value for $\beta$ , the algorithm ensures a balanced and responsive adaptation to the accumulating target data. Such a setup allows our ensemble model to systematically transition its focus, increasingly prioritizing the most recent data, which is presumed to be the most relevant for the current learning context. This approach exemplifies the nuanced and progressive nature of online learning, where the model dynamically adapts to the evolving data landscape, ensuring that its classifications remain accurate and pertinent. V-B Results and Analysis Our primary objective is to develop a model adept at classifying data from the current season, which is data from the season 2019-2020 in this study. To achieve this, we trained individual models for each season ranging from 2011 to 2018. Subsequently, we evaluated the performance of these models using all of the data from season 2019-2020 to ascertain their effectiveness in handling present-season data. The results of using these hyperparameters are presented in Table III, providing a clear overview of each model’s performance across this temporal span. TABLE III: Results of individual models | Season | AUROC | | --- | --- | | 2011-2012 | 0.800 | | 2012-2013 | 0.628 | | 2013-2014 | 0.637 | | 2014-2015 | 0.765 | | 2015-2016 | 0.838 | | 2016-2017 | 0.867 | | 2017-2018 | 0.866 | | 2018-2019 | 0.865 | For the results, a clear upward trend in AUROC scores is evident as we approach the season of 2019-2020, although 2011-2012 is an outlier. This trend is especially pronounced in models trained with more recent data, indicating an improvement in the classifiers’ ability to distinguish between positive and negative cases as they are trained on data closer to the target season. This improvement can be attributed to the increasing relevance and recency of the data. As the training data gets closer to the target season, it likely becomes more representative of the current patterns and trends, thereby enhancing the classifiers’ accuracy. To evaluate the performance of MSAW strategy, we compare them with baseline methods, including: - Pre-trained Using Source: This approach involves pre-training a model using a combined dataset from all source domains. - Online Learning Model: This model employs an online learning strategy, focusing exclusively on sequentially acquired data from the target domain for training. In this approach, the model is tested on an incoming data point and then trained on it before moving on to the next, thus continuously updating its parameters in a test-and-train cycle to adapt to new data. - Online Learning Model Pre-trained Using Source: This approach combines two previous methods together by pre-training a model using all source domains and online learning on the target domain. - Ensemble Method with Equal Weights: This method employs an ensemble method that integrates various individual source models, assigning them equal and fixed weights throughout the process. - Ensemble Method with Volume-based Weights: This method uses an ensemble method that weights individual source models according to the volume of their respective training datasets. - Ensemble Method with Time-based Weights: This ensemble method integrates individual models based on their temporal proximity to the testing season (target domain), defining ‘time distance’ as the number of seasons separating a model’s training data from the testing period. In our comparative analysis of various online learning strategies against the backdrop of sequential epidemiological data, distinct performance metrics emerge, highlighting the efficacy and adaptability of these models as shown in Table IV. The model pre-trained using all of the source domains, at an AUROC of 0.852, demonstrates the potential advantages of a comprehensive, cross-domain dataset for pre-training, indicating the value of extensive historical data. The online learning model, with an AUROC of 0.850, showcases a commendable capacity for leveraging sequentially acquired data, underscoring the viability of online learning in dynamically evolving domains. Online learning started with a pre-trained model boosts the AUROC to 0.864. Conversely, the ensemble with equal weights method registers a lower AUROC of 0.728, suggesting that uniform weighting across diverse models may dilute the classification precision. More notably, the ensemble with volume-based weights, at an AUROC of 0.680, indicates that volume-centric model weighting may not effectively capture the intricate data nuances necessary for accurate target domain classifications. An improvement is observed with the ensemble with time-based weights, achieving an AUROC of 0.761, which implies that temporal relevance plays a crucial role in enhancing model performance, albeit not to the extent of the more sophisticated methods. Our MSAW strategy exhibits an AUROC of 0.870, which is superior to all other methods. This delineates the MSAW strategy’s predominant advantage in dynamically calibrating model weights based on the evolving data landscape, marking a significant stride towards optimizing predictive models for healthcare applications, particularly in the nuanced domain of RSV data analysis. TABLE IV: Comparative results of different models, featuring P values from DeLong’s test for statistical differences between one AUROC and the AUROC of the MSAW method. | Method | AUROC | | --- | --- | | Pre-trained Using Source | 0.852 (P $<$ 0.001) | | Online Learning | 0.850 (P=0.037) | | Online Learning Model Pre-trained Using Source | 0.864 (P $<$ 0.001) | | Ensemble with Equal Weights | 0.728 (P $<$ 0.001) | | Ensemble with Volume-based Weights | 0.680 (P $<$ 0.001) | | Ensemble with Time-based Weights | 0.761 (P $<$ 0.001) | | MSAW | 0.870 | <details> <summary>2402.01987v2/x2.png Details</summary>  ### Visual Description # Technical Document Extraction ## Chart Description The image contains two scatter plots (subplots **a** and **b**) analyzing feature probabilities related to RSV (Respiratory Syncytial Virus) across different years and biological features. Both plots use color-coded data points to represent temporal and categorical trends. --- ### **Subplot a: P(feature|RSV) Over Time** #### **Axis Labels** - **Y-axis**: `P(feature|RSV)` (Probability of feature given RSV) - **X-axis**: Features (e.g., `AGE - less5`, `Platelets - H`, `Eosinophils - L`, etc.) #### **Legend** - **Location**: Right side of the plot - **Categories**: Years (2011–2020) with color codes: - `2011to2012`: Pink - `2012to2013`: Red - `2013to2014`: Light pink - `2014to2015`: Orange - `2015to2016`: Green - `2016to2017`: Dark green - `2017to2018`: Light blue - `2018to2019`: Blue - `2019to2020`: Purple #### **Key Trends** 1. **AGE - less5**: - Highest P(feature|RSV) in `2011to2012` (~0.95). - Gradual decline to ~0.7 by `2019to2020`. 2. **Platelets - H**: - Peaks at ~0.2 in `2011to2012`, drops to ~0.1 by `2019to2020`. 3. **Eosinophils - L**: - Rises from ~0.1 in `2011to2012` to ~0.2 in `2016to2017`, then declines. 4. **Monocytes/100 leukocytes - H**: - Stable (~0.1–0.2) across all years. 5. **Anion gap - L**: - Peaks at ~0.2 in `2017to2018`, then decreases. 6. **Potassium - H**: - Peaks at ~0.15 in `2014to2015`, then declines. 7. **Oxygen saturation - N/L**: - Stable (~0.05–0.1) across years. 8. **C reactive protein - N/H**: - Peaks at ~0.15 in `2018to2019`, then declines. 9. **Magnesium - H**: - Peaks at ~0.15 in `2018to2019`, then declines. 10. **Increased work of breathing**: - Peaks at ~0.15 in `2018to2019`, then declines. --- ### **Subplot b: Log10(P(feature|RSV)/P(feature|non-RSV))** #### **Axis Labels** - **Y-axis**: `Log10(P(feature|RSV)/P(feature|non-RSV))` - **X-axis**: Features (same as subplot a, with suffixes `-H`, `-N`, `-L`). #### **Legend** - **Location**: Below subplot b - **Categories**: - `-H`: High (e.g., `Platelets - H`) - `-N`: Normal (e.g., `Oxygen saturation - N`) - `-L`: Low (e.g., `Eosinophils - L`) #### **Key Trends** 1. **AGE - less5**: - Highest ratio (~3.0) in `2011to2012`, declines to ~0.8 by `2019to2020`. 2. **Platelets - H**: - Peaks at ~2.5 in `2011to2012`, drops to ~1.0 by `2019to2020`. 3. **Eosinophils - L**: - Peaks at ~1.5 in `2014to2015`, declines to ~0.5 by `2019to2020`. 4. **Monocytes/100 leukocytes - H**: - Stable (~0.5–1.0) across years. 5. **Anion gap - L**: - Peaks at ~2.0 in `2013to2014`, declines to ~0.5 by `2019to2020`. 6. **Potassium - H**: - Peaks at ~1.5 in `2014to2015`, declines to ~0.5 by `2019to2020`. 7. **Oxygen saturation - N/L**: - Stable (~0.5–1.0) across years. 8. **C reactive protein - N/H**: - Peaks at ~2.5 in `2013to2014`, declines to ~1.0 by `2019to2020`. 9. **Magnesium - H**: - Peaks at ~2.0 in `2013to2014`, declines to ~1.0 by `2019to2020`. 10. **Increased work of breathing**: - Peaks at ~1.5 in `2018to2019`, declines to ~0.5 by `2019to2020`. --- ### **Cross-Referenced Observations** - **Color Consistency**: - Subplot a’s year colors match subplot b’s suffix colors (e.g., `2011to2012` pink corresponds to `-H` in subplot b). - **Temporal Decline**: Most features show decreasing P(feature|RSV) and Log10 ratios over time, suggesting reduced RSV-associated feature prevalence. - **Feature-Specific Patterns**: - `AGE - less5` and `Platelets - H` exhibit the steepest declines. - Features like `Oxygen saturation - N/L` remain stable. --- ### **Conclusion** The plots highlight temporal trends in RSV-associated features, with most showing declining probabilities over the decade. Subplot b emphasizes the relative importance of features (via Log10 ratios), where `-H` (high) features dominate early years but diminish later. </details> Figure 2: (a) Probability of a feature present given RSV (b) $Log_{10}LR$ of features present. Our proposed MSAW approach, which utilizes dynamically updated weights, surpasses these conventional baseline methods. The essence of its enhanced performance lies in its ability to dynamically adapt, customizing its relevance to the current data. This dynamic adaptation is achieved by continuously adjusting the weights of the models based on their immediate performance with new data. Such a strategy enables the model to respond effectively to changing data trends and to identify patterns that were not apparent in earlier training datasets. As a result, our method shows an improvement in both accuracy and effectiveness in classifying medical data, outperforming the other methodologies evaluated in the study. Fig. 2 (a) presents a comparison of the probabilities of features present in RSV cases across different seasons, as calculated in our training dataset. Fig. 2 (b) compares the $Log_{10}$ likelihood ratios (LR) between RSV cases and non-RSV encounters. A positive LR value suggests that the presence of the feature is more probable in RSV cases than in non-RSV encounters. Among the features for which the probability P(feature $|$ RSV) exceeded 0.01 in the 2019-2020 season, we identified the top 13 features with the LR. In Fig. 2 (a), the probability of ‘Age-less than 5’ group appears high with values ranging from 0.68 to 0.96 in all years, while other features only have around or below 0.2. As shown in Fig. 2 (b), with relatively small ranges of values, ‘Age-less than 5’ and ‘Increased work of breathing’ are two significantly steadier features than others observed through multiple seasons. The values in models of seasons 2012-2013 and 2013-2014 are the most divergent from the model of 2019-2020 season, suggesting that the feature patterns in the 2012-2013 and 2013-2014 seasons are the least similar to our target season in 2019-2020, which also aligns with our model results in Table III where seasons 2012-2013 and 2013-2014 present the lowest AUROC. Conversely, the values for models corresponding to the 2017-2018 and 2018-2019 seasons are predominantly found nearest to the 2019-2020 season, and their respective models also demonstrate strong performance, as detailed in Table III. VI Discussion In this study, we have analyzed various ensemble methods applied to the challenging task of medical data classification. The primary goal was to develop a robust classifier for the season of 2019-2020, leveraging historical data spanning from 2011 to 2018. The results obtained from our experiments do not necessarily reveal a temporal trend in the performance of individual models. The model developed using data from season 2011-2012 performed much better than season 2012-2013, 2013-2014, and 2014-2015, highlighting the importance of allowing the model to autonomously discern relevant patterns and relationships. This finding emphasizes the importance of adopting dynamic and data-driven approaches in medical data classification, as the temporal relevance of information may not always align with conventional intuitions. The evaluation of ensemble methods against baseline models provided valuable insights. Although the AUROCs of the ‘Online Learning Model Pre-trained Using Source’ method and our proposed MSAW are comparable, the MSAW is performing slightly better and is more adaptable to situations in which recent data experience a major domain shift. The MSAW might be expected to perform significantly better when there is a large domain shift in the recent past, for example, new strains of RSV that are symptomatically quite different than previous strains. The MSAW method outperformed other techniques. This superior performance can be attributed to the method’s inherent adaptability, which allows it to continuously learn and assign weights to individual models based on their performance on the most recent data. By doing so, our approach effectively captures emerging patterns and evolving relationships within the data, resulting in enhanced classification accuracy. In an environment where medical knowledge and practice are continually evolving, our research underscores the potential of dynamic ensemble methods to provide more accurate and efficient solutions for medical classification tasks. These findings have significant implications for healthcare decision-making, as they demonstrate the importance of data-driven adaptability in capturing the intricate nuances of medical data. As shown in Fig. 2, many laboratory tests are correlated with RSV presence. For instance, there is a notable correlation between higher platelet counts and the presence of RSV. This correlation is also mentioned in [46], which has shown that children diagnosed with RSV tend to have significantly higher platelet counts compared to those with negative respiratory virus panels or infections caused by other respiratory viruses, including HMPV, influenza A, influenza B, Adenovirus, and Parainfluenza. As another example, our models indicate that patients with RSV often exhibit higher Monocytes counts per 100 leukocytes. One research [47] aligns with our observation, showing that RSV may trigger monocyte activation. Interestingly, additional findings reveal that even normal levels of oxygen saturation and C-reactive protein are associated with RSV. The act of performing these tests, as opposed to not performing these tests, shows a relationship with RSV presence. All these findings highlight the effectiveness of using observational EHRs to detect correlations between lab results and RSV. Importantly, Fig. 2 demonstrates annual variations in the correlation strength between laboratory results and RSV, possibly due in part to adjustments in laboratory testing protocols for different age groups and patient severity levels. In spite of the promising results presented in this study, there remain several noteworthy limitations that warrant attention in future research. In our study, while the Naive Bayes models proved effective, they represent only one simple modeling technique for medical data classification. For more intricate classification tasks, the Naive Bayes method may not effectively capture the complexities of medical data as well as more advanced algorithms, such as neural networks. Therefore, future research should explore incorporating these advanced machine learning techniques, which may offer an improvement in the modeling of the nuanced patterns in medical data. Additionally, the dynamic penalty factor used in our study, though effective, has room for refinement. For example, enhancing its adaptability to the changing data distributions in the target domain could lead to more accurate and responsive models. Specifically, rather than uniformly penalizing all weights, we aim to explore more sophisticated methods to adjust the weights of the model’s parameters for different features to varying extents. This approach would enable a more accurate reflection of changes. Another limitation of our study stems from our dataset. RSV cases are identified using ICD codes rather than confirmed through RSV testing results. Consequently, non-RSV encounters might include some RSV-positive cases lacking RSV-related ICD codes. However, given the vast number of non-RSV encounters, this issue is unlikely to significantly impact the results. In addition, the specificity of our dataset, which was centered around RSV information from UPMC, presents a potential constraint in terms of the generalizability of our findings across different medical contexts. The healthcare field is diverse, and patterns observed in one dataset may not be universally applicable. Our future research will aim to validate the model across a broader spectrum of biomedical scenarios. For example, we can incorporate data from different medical institutions, which would help assess the transferability of our approach and determine necessary modifications for different settings. Expanding the scope of our research through collaborative and cross-disciplinary studies could also provide insights and data sources, enhancing the model’s applicability and robustness. VII Conclusion This paper introduces an online multi-source transfer learning methodology featuring a dynamic weighting mechanism, applied to the development of RSV case detection models using multiple years’ EHR data from the UPMC. Experimental results demonstrate that this method significantly outperforms many baseline techniques, highlighting the considerable potential of the approach. Notably, it adeptly addresses the challenges inherent in biomedical research, such as the scarcity of labeled data and the complexities associated with sequential data. The results of our investigation provide support for the effectiveness of our MSAW method, but also pave the way for studying more advanced classification models for sequential data analysis. VIII Acknowledgements This work was supported by the research grants R00LM013383 and R01LM013509 from the National Library of Medicine, National Institutes of Health (NIH). The research data were retrieved by the University of Pittsburgh’s Health Record Research Request, which was supported by the NIH grant UL1TR001857. Dr. Harry Hochheiser was supported by MIDAS grant U24GM132013. Dr. Jessi Espino was partially supported by the National Center for Immunization and Respiratory Diseases grant 5U01IP001184. We are grateful for the invaluable assistance and guidance from Dr. Richard K. Zimmerman and Dr. Mary Patricia Nowalk, and Dr. Milos Hauskrecht, from the University of Pittsburgh. References - [1] S. Sukhija, N. C. Krishnan, and G. Singh, “Supervised heterogeneous domain adaptation via random forests.” in Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence (IJCAI-16), 2016, pp. 2039–2045. - [2] H. Wu, Q. Wu, and M. K. Ng, “Knowledge preserving and distribution alignment for heterogeneous domain adaptation,” ACM Transactions on Information Systems (TOIS), vol. 40, no. 1, pp. 1–29, 2021. - [3] J. T. Zhou, I. W. Tsang, S. J. Pan, and M. Tan, “Multi-class heterogeneous domain adaptation,” Journal of Machine Learning Research, 2019. - [4] P. Zhao, S. C. Hoi, J. Wang, and B. Li, “Online transfer learning,” Artificial intelligence, vol. 216, pp. 76–102, 2014. - [5] Z. Kang, B. Yang, S. Yang, X. Fang, and C. Zhao, “Online transfer learning with multiple source domains for multi-class classification,” Knowledge-Based Systems, vol. 190, p. 105149, 2020. - [6] P. Zhao and S. C. Hoi, “Otl: A framework of online transfer learning.(2010),” in Proceedings of the 27th International Conference on Machine Learning (ICML 2010): Haifa, Israel, 21-24 June, pp. 219–1. - [7] S. Agrahari and A. K. Singh, “Concept drift detection in data stream mining: A literature review,” Journal of King Saud University-Computer and Information Sciences, vol. 34, no. 10, pp. 9523–9540, 2022. - [8] G. Piedimonte and M. K. Perez, “Respiratory syncytial virus infection and bronchiolitis,” Pediatrics in review, vol. 35, no. 12, pp. 519–530, 2014. - [9] T. Shi, D. A. McAllister, K. L. O’Brien, E. A. Simoes, S. A. Madhi, B. D. Gessner, F. P. Polack, E. Balsells, S. Acacio, C. Aguayo et al., “Global, regional, and national disease burden estimates of acute lower respiratory infections due to respiratory syncytial virus in young children in 2015: a systematic review and modelling study,” The Lancet, vol. 390, no. 10098, pp. 946–958, 2017. - [10] C. B. Hall, G. A. Weinberg, M. K. Iwane, A. K. Blumkin, K. M. Edwards, M. A. Staat, P. Auinger, M. R. Griffin, K. A. Poehling, D. Erdman, C. G. Grijalva, Y. Zhu, and P. Szilagyi, “The burden of respiratory syncytial virus infection in young children,” New England Journal of Medicine, vol. 360, no. 6, pp. 588–598, 2009. - [11] A. R. Falsey, P. A. Hennessey, M. A. Formica, C. Cox, and E. E. Walsh, “Respiratory syncytial virus infection in elderly and high-risk adults,” N Engl J Med, vol. 352, no. 17, pp. 1749–59, 2005, 1533-4406 Falsey, Ann R Hennessey, Patricia A Formica, Maria A Cox, Christopher Walsh, Edward E R01-AI-45969/AI/NIAID NIH HHS/United States Journal Article Research Support, U.S. Gov’t, P.H.S. United States 2005/04/29 N Engl J Med. 2005 Apr 28;352(17):1749-59. doi: 10.1056/NEJMoa043951. - [12] C. Griffiths, S. J. Drews, and D. J. Marchant, “Respiratory syncytial virus: infection, detection, and new options for prevention and treatment,” Clinical microbiology reviews, vol. 30, no. 1, pp. 277–319, 2017. - [13] F. Zhuang, Z. Qi, K. Duan, D. Xi, Y. Zhu, H. Zhu, H. Xiong, and Q. He, “A comprehensive survey on transfer learning,” Proceedings of the IEEE, vol. 109, no. 1, pp. 43–76, 2020. - [14] K. Weiss, T. M. Khoshgoftaar, and D. Wang, “A survey of transfer learning,” Journal of Big data, vol. 3, no. 1, pp. 1–40, 2016. - [15] L. Zhang and X. Gao, “Transfer adaptation learning: A decade survey,” IEEE Transactions on Neural Networks and Learning Systems, 2022. - [16] N. Agarwal, A. Sondhi, K. Chopra, and G. Singh, “Transfer learning: Survey and classification,” in Smart Innovations in Communication and Computational Sciences, S. Tiwari, M. C. Trivedi, K. K. Mishra, A. Misra, K. K. Kumar, and E. Suryani, Eds. Singapore: Springer Singapore, 2021, pp. 145–155. - [17] R. Bao, Y. Sun, Y. Gao, J. Wang, Q. Yang, H. Chen, Z.-H. Mao, and Y. Ye, “A survey of heterogeneous transfer learning,” arXiv preprint arXiv:2310.08459, 2023. - [18] C. Tan, F. Sun, T. Kong, W. Zhang, C. Yang, and C. Liu, “A survey on deep transfer learning,” in Artificial Neural Networks and Machine Learning–ICANN 2018: 27th International Conference on Artificial Neural Networks, Rhodes, Greece, October 4-7, 2018, Proceedings, Part III 27. Springer, 2018, pp. 270–279. - [19] M. Iman, H. R. Arabnia, and K. Rasheed, “A review of deep transfer learning and recent advancements,” Technologies, vol. 11, no. 2, p. 40, 2023. - [20] H. Liang, W. Fu, and F. Yi, “A survey of recent advances in transfer learning,” in 2019 IEEE 19th international conference on communication technology (ICCT). IEEE, 2019, pp. 1516–1523. - [21] Z. Zhu, K. Lin, A. K. Jain, and J. Zhou, “Transfer learning in deep reinforcement learning: A survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence, pp. 1–20, 2023. - [22] S. Saha and T. Ahmad, “Federated transfer learning: Concept and applications,” Intelligenza Artificiale, vol. 15, no. 1, pp. 35–44, 2021. - [23] E. Hallaji, R. Razavi-Far, and M. Saif, “Federated and transfer learning: A survey on adversaries and defense mechanisms,” in Federated and Transfer Learning. Springer, 2022, pp. 29–55. - [24] M. Wang and W. Deng, “Deep visual domain adaptation: A survey,” Neurocomputing, vol. 312, pp. 135–153, 2018. - [25] J. Zhang, B. Wang, Z. Hu, P. W. W. Koh, and A. J. Ratner, “On the trade-off of intra-/inter-class diversity for supervised pre-training,” Advances in Neural Information Processing Systems, vol. 36, 2024. - [26] Z. Alyafeai, M. S. AlShaibani, and I. Ahmad, “A survey on transfer learning in natural language processing,” arXiv preprint arXiv:2007.04239, 2020. - [27] R. Liu, Y. Shi, C. Ji, and M. Jia, “A survey of sentiment analysis based on transfer learning,” IEEE access, vol. 7, pp. 85 401–85 412, 2019. - [28] S. Ruder, M. E. Peters, S. Swayamdipta, and T. Wolf, “Transfer learning in natural language processing,” in Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: Tutorials, 2019, pp. 15–18. - [29] X. Yu, J. Wang, Q.-Q. Hong, R. Teku, S.-H. Wang, and Y.-D. Zhang, “Transfer learning for medical images analyses: A survey,” Neurocomputing, vol. 489, pp. 230–254, 2022. - [30] A. Sufian, A. Ghosh, A. S. Sadiq, and F. Smarandache, “A survey on deep transfer learning to edge computing for mitigating the COVID-19 pandemic,” Journal of Systems Architecture, vol. 108, p. 101830, 2020. - [31] S. Wei, R. Moore, H. Zhang, Y. Xie, and R. Kamaleswaran, “Transfer causal learning: Causal effect estimation with knowledge transfer,” in ICML 3rd Workshop on Interpretable Machine Learning in Healthcare (IMLH), 2023. - [32] S. Wei, X. Kong, S. A. Huestis-Mitchell, Y. Xie, S. Zhu, A. S. Xavier, and F. Qiu, “Unfairness detection within power systems through transfer counterfactual learning,” in Causal Representation Learning Workshop at NeurIPS 2023, 2023. - [33] Z. Lai, J. Wu, S. Chen, Y. Zhou, A. Hovakimyan, and N. Hovakimyan, “Language models are free boosters for biomedical imaging tasks,” arXiv preprint arXiv:2403.17343, 2024. - [34] L. J. Wong and A. J. Michaels, “Transfer learning for radio frequency machine learning: A taxonomy and survey,” Sensors, vol. 22, no. 4, p. 1416, 2022. - [35] N. Maray, A. H. Ngu, J. Ni, M. Debnath, and L. Wang, “Transfer learning on small datasets for improved fall detection,” Sensors, vol. 23, no. 3, p. 1105, 2023. - [36] H. Wu, Y. Yan, Y. Ye, H. Min, M. K. Ng, and Q. Wu, “Online heterogeneous transfer learning by knowledge transition,” ACM Transactions on Intelligent Systems and Technology (TIST), vol. 10, no. 3, pp. 1–19, 2019. - [37] Z. Lai, X. Zhang, and S. Chen, “Adaptive ensembles of fine-tuned transformers for llm-generated text detection,” arXiv preprint arXiv:2403.13335, 2024. - [38] Z. Hu, J. Zhang, H. Wang, S. Liu, and S. Liang, “Leveraging relational graph neural network for transductive model ensemble,” in Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2023, pp. 775–787. - [39] Y. Yan, Q. Wu, M. Tan, M. K. Ng, H. Min, and I. W. Tsang, “Online heterogeneous transfer by hedge ensemble of offline and online decisions,” IEEE transactions on neural networks and learning systems, vol. 29, no. 7, pp. 3252–3263, 2017. - [40] C. F. Tso, C. Lam, J. Calvert, and Q. Mao, “Machine learning early prediction of respiratory syncytial virus in pediatric hospitalized patients,” Frontiers in Pediatrics, vol. 10, p. 886212, 2022. - [41] M. Cord—De Iaco, A. Kimberly, F. Gesualdo, E. Pandolfi, I. Croci, and A. E. Tozzi, “Machine learning clinical decision support systems for surveillance: a case study on pertussis and rsv in children,” Frontiers in Pediatrics, vol. 11, p. 1112074, 2023. - [42] Y. Ji, Y. Gao, R. Bao, Q. Li, D. Liu, Y. Sun, and Y. Ye, “Prediction of covid-19 patients’ emergency room revisit using multi-source transfer learning,” in 2023 IEEE 11th International Conference on Healthcare Informatics (ICHI). IEEE, 2023, pp. 138–144. - [43] S. Visweswaran, B. McLay, N. Cappella, M. Morris, J. T. Milnes, S. E. Reis, J. C. Silverstein, and M. J. Becich, “An atomic approach to the design and implementation of a research data warehouse,” Journal of the American Medical Informatics Association, vol. 29, no. 4, pp. 601–608, 2022. - [44] A. R. Aronson and F.-M. Lang, “An overview of metamap: historical perspective and recent advances,” Journal of the American Medical Informatics Association, vol. 17, no. 3, pp. 229–236, 2010. - [45] A. L. Pineda, Y. Ye, S. Visweswaran, G. F. Cooper, M. M. Wagner, and F. R. Tsui, “Comparison of machine learning classifiers for influenza detection from emergency department free-text reports,” Journal of biomedical informatics, vol. 58, pp. 60–69, 2015. - [46] S. Elber-Dorozko, L. Kerem, D. Wolf, S. Brodie, Y. Berkun, R. Brooks, and O. Breuer, “Platelet count and risk of severe illness in hospitalized children with influenza-like illness,” Acta Paediatrica, 2023. - [47] J. L. Kimpen, “Respiratory syncytial virus and asthma: the role of monocytes,” American journal of respiratory and critical care medicine, vol. 163, no. supplement_1, pp. S7–S9, 2001.