External Data Sources Research Articles

Using real-world data to predict health outcomes-The prediction design: Application and sample size planning.

The gold standard for investigating the efficacy of a new therapy is a (pragmatic) randomized controlled trial (RCT). This approach is costly, time-consuming, and not always practicable. At the same time, huge quantities of available patient-level control condition data in analyzable format of (former) RCTs or real-world data (RWD) are neglected. Therefore, alternative study designs are desirable. The design presented here consists of setting up a prediction model for determining treatment effects under the control condition for future patients. When a new treatment is intended to be tested against a control treatment, a single-arm trial for the new therapy is conducted. The treatment effect is then evaluated by comparing the outcomes of the single-arm trial against the predicted outcomes under the control condition. While there are obvious advantages of this design compared to classical RCTs (increased efficiency, lower cost, alleviating participants' fear of being on control treatment), there are several sources of bias. Our aim is to investigate whether and how such a design-the prediction design-may be used to provide information on treatment effects by leveraging external data sources. For this purpose, we investigated under what assumptions linear prediction models could be used to predict the counterfactual of patients precisely enough to construct a test and an appropriate sample size formula for evaluating the average treatment effect in the population of a new study. A user-friendly R Shiny application (available at: https://web.imbi.uni-heidelberg.de/PredictionDesignR/) facilitates the application of the proposed methods, while a real-world application example illustrates them.

S2MetNet: A novel dataset and deep learning benchmark for methane point source quantification using Sentinel-2 satellite imagery

Methane, as a crucial greenhouse gas, plays a significant role in global warming, contributing to approximately one-quarter of the observed climate change since pre-industrial times. Consequently, the detection and quantification of major methane emitters are vital in addressing this issue effectively. Satellite sensors with shortwave infrared (SWIR) spectral bands provide valuable information for monitoring methane emissions. For example, Sentinel-2 multispectral data have the capability to detect methane plumes of large point sources. As such, a wide range of quantification approaches have been developed to quantify methane source rates based on this dataset. Most of the existing methods, however, require auxiliary data, such as wind speed, and have large uncertainties. In this study, we introduce a novel approach based on deep learning models to enhance the precision of methane quantification using Sentinel-2 data without the reliance on external data sources. To train the proposed deep learning model, a comprehensive benchmark dataset has been generated, using Sentinel-2 data. This dataset is created by integrating simulated plumes and background noise extracted from real Sentinel-2 images. This approach ensures the integration of realistic environmental conditions within the simulated data, enhancing the robustness and reliability of our proposed model. The generated benchmark dataset is utilized in different deep learning architectures, namely VGG-19, ResNet-50, Inception-v3, DenseNet-121, Swin-T, and EfficientNet-V2L, to estimate methane source rate. The performance of deep models has been evaluated in three learning strategies, namely from scratch, transfer-learning, and fine-tuning. The fine-tuned EfficientNet-V2L achieves the highest accuracy with root-mean-square error (RMSE), mean absolute percentage error (MAPE), and Pearson R of 2101 kg h−1, 10.05%, and 95.70%, respectively. More importantly, the proposed model demonstrates superior performance compared to conventional physical-based quantification methods (e.g., integrated mass enhancement) and recently developed deep learning model techniques (e.g., MethaNet). In particular, the proposed model exhibits an improvement of approximately 1287 kg h−1 in terms of RMSE, a 3.92% reduction in MAPE, and a 5.01% enhancement in R compared to the IME method. These results highlight the advancements achieved by the proposed approach in accurately quantifying methane emissions using Sentinel-2 imagery. The generated benchmark dataset and the developed deep learning model presented in this study serve as a fundamental resource and constructive framework for future research, promoting extensive implementation across various methane monitoring scenarios on different satellites and in distinct geographic regions, which delivering greater effectiveness to global methane emission monitoring initiatives.

An AI Approach to Generating MIDD Assets Across the Drug Development Continuum

Model-informed drug development involves developing and applying exposure-based, biological, and statistical models derived from preclinical and clinical data sources to inform drug development and decision-making. Discrete models are generated from individual experiments resulting in a single model expression that is utilized to inform a single stage-gate decision. Other model types provide a more holistic view of disease biology and potentially disease progression depending on the appropriateness of the underlying data sources for that purpose. Despite this awareness, most data integration and model development approaches are still reliant on internal (within company) data stores and traditional structural model types. An AI/ML-based MIDD approach relies on more diverse data and is informed by past successes and failures including data outside a host company (external data sources) that may enhance predictive value and enhance data generated by the sponsor to reflect more informed and timely experimentation. The AI/ML methodology also provides a complementary approach to more traditional modeling efforts that support MIDD and thus yields greater fidelity in decision-making. Early pilot studies support this assessment but will require broader adoption and regulatory support for more evidence and refinement of this paradigm. An AI/ML-based approach to MIDD has the potential to transform regulatory science and the current drug development paradigm, optimize information value, and increase candidate and eventually product confidence with respect to safety and efficacy. We highlight early experiences with this approach using the AI compute platforms as representative examples of how MIDD can be facilitated with an AI/ML approach.Graphical abstract

Effective Entity Augmentation by Querying External Data Sources

Users often want to augment and enrich entities in their datasets with relevant information from external data sources. As many external sources are accessible only via keyword-search interfaces, a user usually has to manually formulate a keyword query that extract relevant information for each entity. This approach is challenging as many data sources contain numerous tuples, only a small fraction of which may contain entity-relevant information. Furthermore, different datasets may represent the same information in distinct forms and under different terms (e.g., different data source may use different names to refer to the same person). In such cases, it is difficult to formulate a query that precisely retrieves information relevant to an entity. Current methods for information enrichment mainly rely on lengthy and resource-intensive manual effort to formulate queries to discover relevant information. However, in increasingly many settings, it is important for users to get initial answers quickly and without substantial investment in resources (such as human attention). We propose a progressive approach to discovering entity-relevant information from external sources with minimal expert intervention. It leverages end users' feedback to progressively learn how to retrieve information relevant to each entity in a dataset from external data sources. Our empirical evaluation shows that our approach learns accurate strategies to deliver relevant information quickly.

EPPO ontology: a semantic-driven approach for plant and pest codes representation.

The agricultural industry and regulatory organizations define strategies and build tools and products for plant protection against pests. To identify different plants and their related pests and avoid inconsistencies between such organizations, an agreed and shared classification is necessary. In this regard, the European and Mediterranean Plant Protection Organization (EPPO) has been working on defining and maintaining a harmonized coding system (EPPO codes). EPPO codes are an easy way of referring to a specific organism by means of short 5 or 6 letter codes instead of long scientific names or ambiguous common names. EPPO codes are freely available in different formats through the EPPO Global Database platform and are implemented as a worldwide standard and used among scientists and experts in both industry and regulatory organizations. One of the large companies that adopted such codes is BASF, which uses them mainly in research and development to build their crop protection and seeds products. However, extracting the information is limited by fixed API calls or files that require additional processing steps. Facing these issues makes it difficult to use the available information flexibly, infer new data connections, or enrich it with external data sources. To overcome such limitations, BASF has developed an internal EPPO ontology to represent the list of codes provided by the EPPO Global Database as well as the regulatory categorization and relationship among them. This paper presents the development process of this ontology along with its enrichment process, which allows the reuse of relevant information available in an external knowledge source such as the NCBI Taxon. In addition, this paper describes the use and adoption of the EPPO ontology within the BASF's Agricultural Solutions division and the lessons learned during this work.

AI in Cybersecurity: The Paradox

Modern artificial intelligence is inherently paradoxical in many ways. While AI aims to increase automation, it also requires more intimate human involvement to reflect on the insights generated (automation paradox). While AI results in job displacement, it also creates new jobs, some simply to provide the necessary support systems for those newly unemployed (transition paradox). And as generative AI takes away control over the creative process, it also offers new creative opportunities (creativity paradox). This article considers another paradox, that relates to the fact that computational systems created using AI can be used both for public good in civilian applications and for harm across a range of application areas and settings [A1]. This contradiction is explored within an organizational and governmental context, where modern AI relies on data which might be externally or internally-sourced [A2]. External data sources [A3] are inclusive of open-source intelligence (OS-INT), such as information available on the Internet and the dark web, and internal data sources may include proprietary data found within an organizational or a wider governmental context [A4]. A further relevant consideration is the expanding role of the Internet of Things to support smart infrastructures, which has created new vulnerabilities [A5].

A trial-based cost-utility analysis of sugemalimab vs. placebo as consolidation therapy for unresectable stage III NSCLC in China.

The effectiveness of sugemalimab vs. placebo in post-chemoradiotherapy patients with locally advanced, unresectable stage III NSCLC has been demonstrated and approved by China National Medical Products Administration. The purpose of this study was to assess the cost-effectiveness of sugemalimab vs. placebo for consolidation treatment of stage III NSCLC from the perspective of the Chinese healthcare system. A 3-state Markov model with a 3-week cycle length was performed to appraise the incremental cost-utility ratio (ICUR) of sugemalimab consolidation therapy based on the GEMSTONE-301 clinical trial over a 10-year time horizon. Only direct medical costs, including costs of drug (maintenance and subsequent treatment), routine follow-up, best supportive care, and terminal care in end of life were considered in this model. Costs and health utilities were obtained from local databases and published articles. Sensitivity and scenario analyses were adopted to evaluate the model uncertainty. Internal and external data sources were used to justify the plausibility of the extrapolated portion of the survival model chosen. In comparison with the placebo, sugemalimab consolidation therapy was not cost-effective as it yielded an ICUR value of $90,277 and $49,692 for the concurrent chemoradiotherapy (cCRT) and the sequential chemoradiotherapy (sCRT) population at the willingness-to-pay (WTP) threshold of $37,663/QALYs, respectively. When taking the sugemalimab patient assistance program (PAP) into consideration, sugemalimab consolidation therapy was cost-effective with an ICUR dramatic decreases below the WTP. Sensitivity analyses demonstrated that the ICUR was most sensitive to the discount rate and subsequent treatment. However, none of the sensitive parameters could affect the cost-effective conclusions without or with PAP. Scenario analyses revealed that the model was particularly affected by assumptions regarding discount in sugemalimab, time horizon, mean duration of sugemalimab maintenance treatment. From the perspective of Chinese healthcare system, sugemalimab consolidation therapy was not a cost-effective strategy in cCRT and sCRT patients with unresectable stage III NSCLC. Given that the sugemalimab PAP was available, sugemalimab consolidation therapy became a cost-effective option.

Using item response theory to develop a shortened practice environment scale of the nursing work index.

The 31-item Practice Environment Scale of the Nursing Work Index (PES-NWI) has been frequently used globally to measure the nurse work environment. However, due to its length and subsequent respondent burden, a more parsimonious version of the PES-NWI may be desirable. Item response theory (IRT) is a statistical technique that assists in decreasing the number of items in an instrument without sacrificing reliability and validity. Two separate samples of nurses in the United States (one called the "internal data source"and the other called "external data source"; sample sizes = 843 and 722, respectively) were analyzed. The internal data source was randomly split into training (n = 531) and validating data sets (n = 312), while a separate whole external data source was used as the final validating data set. Using IRT with training data, we removed nine items; two additional items were removed based on recommendations from a previous study. Confirmatory factor analyses supported the validity of the measurement model with the 20-item of PES-NWI in both internal and external validation data sources. The correlations among subscales between 31- and 20-item versions were high magnitude for five subscales in both validation data sets (τ = 0.84-0.89). Ultimately, we identified a 20-item version of the PES-NWI which demonstrated adequate validity and reliability properties while decreasing data collection burden yet maintaining a similar factor structure to the original instrument. Additional research may be necessary to update the items themselves on the PES-NWI.

Use of Immunization Information Systems in Ascertainment of COVID-19 Vaccinations for Claims-Based Vaccine Safety and Effectiveness Studies

Safety and effectiveness studies of COVID-19 vaccines are being conducted using clinical data, including administrative claims. However, claims data only partially capture administered COVID-19 vaccine doses for numerous reasons, such as vaccination at sites that do not generate claims for reimbursement. To evaluate the extent to which Immunization Information Systems (IIS) data linked to claims data enhances claims-based COVID-19 vaccine capture for a commercially insured population and to estimate the magnitude of misclassification of vaccinated individuals as having unvaccinated status in the linked IIS and claims data. This cohort study used claims data from a commercial health insurance database and obtained vaccination data from IIS repositories in 11 US states. Participants were individuals younger than 65 years who resided in 1 of 11 states of interest and who were insured in health plans from December 1, 2020, through December 31, 2021. Estimated proportion of individuals with at least 1 dose of any COVID-19 vaccine and proportion of individuals with a completed vaccine series based on general population guidelines. Vaccination status estimates were calculated and compared using claims data alone and linked IIS and claims data. Remaining misclassification of vaccination status was assessed by comparing linked IIS and claims data estimates with estimates from external surveillance data sources (Centers for Disease Control and Prevention [CDC] and state Department of Health [DOH]) and capture-recapture analysis. This cohort study included 5 112 722 individuals (mean [SD] age, 33.5 [17.6] years; 2 618 098 females [51.2%]) from 11 states. Characteristics of those who received at least 1 vaccine dose and those who completed a vaccine series were similar to the overall study population. The proportion with at least 1 vaccine dose increased from 32.8% using claims data alone to 48.1% when the data were supplemented with IIS vaccination records. Vaccination estimates using linked IIS and claims data varied widely by state. The percentage of individuals who completed a vaccine series increased from 24.4% to 41.9% after the addition of IIS vaccine records and varied across states. The percentages of underrecording using linked IIS and claims data were 12.1% to 47.1% lower than those using CDC data, 9.1% to 46.9% lower than the state DOH, and 9.2% to 50.9% lower than capture-recapture analysis. Results of this study suggested that supplementing COVID-19 claims records with IIS vaccination records substantially increased the number of individuals who were identified as vaccinated, yet potential underrecording remained. Improvements in reporting vaccination data to IIS infrastructures could allow frequent updates of vaccination status for all individuals and all vaccines.

Generating Small Areal Synthetic Microdata from Public Aggregated Data Using an Optimization Method

Small area microdata contain attributes and locations of individual members of a population in small census geographies. This type of data is critical in research and policymaking, but it is often not publicly available due to confidentiality concerns. The limited access to small area microdata can result in insufficient data for certain research (data scarcity). Even for researchers qualified to access the small area microdata, their research can hardly be reproduced by others (method irreproducibility). To address these issues, we develop a method to generate small area synthetic microdata (SASM) that is suitable for public use. Specifically, an optimization approach is proposed to minimize the difference between published census tables and the SASM. Two counties in Ohio are used as case studies to test the efficacy of the proposed method and the validity of the resulting SASM. The results show that the SASM aligns not only with the census tables, but also with an external data source that contains a sample of the small area microdata. We also illustrate how the SASM can be used to address data scarcity and method irreproducibility in demographic research.

Coping with Information Loss and the Use of Auxiliary Sources of Data: A Report from the NISS Ingram Olkin Forum Series on Unplanned Clinical Trial Disruptions

While the SARS-CoV-2 (COVID-19) pandemic has led to an impressive and unprecedented initiation of clinical research, it has also led to considerable disruption of clinical trials in other disease areas, with around 80% of non-COVID-19 trials stopped or interrupted during the pandemic. In many cases the disrupted trials will not have the planned statistical power necessary to yield interpretable results. This article describes methods to compensate for the information loss arising from trial disruptions by incorporating additional information available from auxiliary data sources. The methods described include the use of auxiliary data on baseline and early outcome data available from the trial itself and frequentist and Bayesian approaches for the incorporation of information from external data sources. The methods are illustrated by application to the analysis of artificial data based on the Primary care pediatrics Learning Activity Nutrition (PLAN) study, a clinical trial assessing a diet and exercise intervention for overweight children, that was affected by the COVID-19 pandemic. We show how all of the methods proposed lead to an increase in precision relative to use of complete case data only.

A Sequence and Network Embedding Method for Bus Arrival Time Prediction Using GPS Trajectory Data Only

Recently, achieving high accuracy in bus arrival time prediction has become a hot research topic. Extensive research has been conducted around this topic, among which fusion of multi-source heterogeneous information is the mainstream. However, using only GPS trajectory data to predict the arrival time of buses is meaningful for small cities since they do not have easy access to external sources of data. Therefore, we propose a novel deep-learning model based on sequence and network(i.e., graph) embedding (SGE-net) which adopts only GPS trajectories as input and achieves accurate bus arrival time prediction without external data. Specifically, the sequence embedding part, composed of the Echo State Network (ESN), the Auto-Encoder (AE), and the k-means cluster method, is constructed to extract the complex and potential sequence patterns from a large number of trajectory sequences, and cluster different patterns of trajectories. Then, the network embedding part with the node2vec algorithm is proposed to capture the spatial correlation among bus stops on different lines under the condition of limited data. Finally, a fusion prediction part based on the specifically designed parallel Gated Recurrent Unit (GRU) networks is introduced to combine sequence and network embedding vectors with hidden outputs of the GRUs and then get the prediction. Both real-world offline and online experiments are conducted based on GPS data of bus lines in QiTaiHe city and ShiJiaZhuang city. The results clearly show that the SGE-net outperforms the baseline models, and validate the effectiveness and robustness of the SGE-net.

Target-based fusion using social determinants of health to enhance suicide prediction with electronic health records.

Preventing suicide in US youth is of paramount concern, with rates increasing over 50% between 2007 and 2018. Statistical modeling using electronic health records may help identify at-risk youth before a suicide attempt. While electronic health records contain diagnostic information, which are known risk factors, they generally lack or poorly document social determinants (e.g., social support), which are also known risk factors. If statistical models are built incorporating not only diagnostic records, but also social determinants measures, additional at-risk youth may be identified before a suicide attempt. Suicide attempts were predicted in hospitalized patients, ages 10-24, from the State of Connecticut's Hospital Inpatient Discharge Database (HIDD; N = 38943). Predictors included demographic information, diagnosis codes, and using a data fusion framework, social determinants features transferred or fused from an external source of survey data, The National Longitudinal Study of Adolescent to Adult Health (Add Health). Social determinant information for each HIDD patient was generated by averaging values from their most similar Add Health individuals (e.g., top 10), based upon matching shared features between datasets (e.g., Pearson's r). Attempts were then modelled using an elastic net logistic regression with both HIDD features and fused Add Health features. The model including fused social determinants outperformed the conventional model (AUC = 0.83 v. 0.82). Sensitivity and positive predictive values at 90 and 95% specificity were almost 10% higher when including fused features (e.g., sensitivity at 90% specificity = 0.48 v. 0.44). Among social determinants variables, the perception that their mother cares and being non-religious appeared particularly important to performance improvement. This proof-of-concept study showed that incorporating social determinants measures from an external survey database could improve prediction of youth suicide risk from clinical data using a data fusion framework. While social determinant data directly from patients might be ideal, estimating these characteristics via data fusion avoids the task of data collection, which is generally time-consuming, expensive, and suffers from non-compliance.

Empowering trusted data sharing for data analytics in a federated environment: A blockchain-based approach

As data analytics is used in business to increase profits, organizations use it to pursue their goals. Even if enterprise data could be already valuable on its own, in many cases, combining it with external data sources would boost the value of the output, making data sharing a need in data analytics. At the same time, organizations are reluctant to share data, as they are scared of disclosing critical information. This calls for solutions that are able to safeguard data holders by regulating how data can be shared to ensure the so-called data sovereignty. This paper focuses on the usage of data lakes as well-established technology across enterprises for data analytics where internal or publicly available data are considered. The goal is to extend data lakes with functionalities that, respecting the data sovereignty, enable a data lake also to be ingested with data shared by other organizations and to share data to external organizations. Notable, the purpose of this work is to face this issue by defining an architecture that, inserted in a federated environment: restricts data access and enables monitoring that the actual usage of data respects the data sovereignty expressed in the policies agreed upon by the involved parties; makes use of Blockchain technology as a means for guaranteeing the traceability of data sharing; and allows for balancing computation movement and data movement. The proposed approach has been applied to a healthcare scenario where several institutions (e.g., hospitals and clinics, research institutes, and medical universities) produce and collect clinical data in local data lakes.

Using AI to detect panic buying and improve products distribution amid pandemic.

The COVID-19 pandemic has triggered panic-buying behavior around the globe. As a result, many essential supplies were consistently out-of-stock at common point-of-sale locations. Even though most retailers were aware of this problem, they were caught off guard and are still lacking the technical capabilities to address this issue. The primary objective of this paper is to develop a framework that can systematically alleviate this issue by leveraging AI models and techniques. We exploit both internal and external data sources and show that using external data enhances the predictability and interpretability of our model. Our data-driven framework can help retailers detect demand anomalies as they occur, allowing them to react strategically. We collaborate with a large retailer and apply our models to three categories of products using a dataset with more than 15million observations. We first show that our proposed anomaly detection model can successfully detect anomalies related to panic buying. We then present a prescriptive analytics simulation tool that can help retailers improve essential product distribution in uncertain times. Using data from the March 2020 panic-buying wave, we show that our prescriptive tool can help retailers increase access to essential products by 56.74%.

Artificial intelligence: what it is and what it can do for dentists

Artificial intelligence (AI) is an increasingly relevant topic for dental clinicians, with AI applications entering the clinical arena at a high pace. This article outlines what AI is, how it works, what its application fields are, but also what challenges the profession faces now and in the future. Computer vision, language processing, simulation and precision dentistry are the main fields where AI is, or will be, applied in dentistry. The ability to be generalizable to external data sources, be accurate, useful and easy to explain are the main cornerstones of AI for health applications. Clinicians should be able to appraise AI applications before integrating them in their daily workflow. AI will be useful for synthesizing an increasing amount of data in dentistry, allowing more automated, efficient and precise care. Certain tools will also facilitate patient communication and documentation. Dentists should critically evaluate AI against certain quality criteria and standards. CPD/Clinical Relevance: It is important to be aware of the applications of artificial intelligence in dentistry.

Streamlining PySpark Data Pipelines: Integrating External Data Sources and APIs

In the ever-expanding domain of big data analytics, integrating external data sources and APIs into data processing workflows is essential for enriching analyses and enhancing decision-making. PySpark, a prominent component of the Apache Spark ecosystem, offers robust capabilities for high-performance data processing but often requires extension to utilize external databases, web services, and third-party APIs effectively. This paper explores diverse methodologies for importing and managing data from these varied sources directly into PySpark. Techniques discussed include utilizing JDBC for database connections, employing HTTP clients for accessing RESTful APIs, and leveraging cloud storage APIs compatible with Hadoop. While integrating external data sources into PySpark pipelines significantly enriches the available data landscape, it also introduces challenges such as scalability of ingestion processes, maintenance of data integrity, and real-time data processing overhead. This study aims to provide a detailed examination of these integration techniques, highlighting both the strategic advantages and potential complications. By addressing these challenges, the paper presents a pathway towards more dynamic and comprehensive data analytics platforms, enabling businesses to leverage real-time insights and drive more nuanced analyses through PySpark.

Midas

Millimeter wave radar is a promising sensing modality for enabling pervasive and privacy-preserving human sensing. However, the lack of large-scale radar datasets limits the potential of training deep learning models to achieve generalization and robustness. To close this gap, we resort to designing a software pipeline that leverages wealthy video repositories to generate synthetic radar data, but it confronts key challenges including i) multipath reflection and attenuation of radar signals among multiple humans, ii) unconvertible generated data leading to poor generality for various applications, and iii) the class-imbalance issue of videos leading to low model stability. To this end, we design Midas to generate realistic, convertible radar data from videos via two components: (i) a data generation network (DG-Net) combines several key modules, depth prediction, human mesh fitting and multi-human reflection model, to simulate the multipath reflection and attenuation of radar signals to output convertible coarse radar data, followed by a Transformer model to generate realistic radar data; (ii) a variant Siamese network (VS-Net) selects key video clips to eliminate data redundancy for addressing the class-imbalance issue. We implement and evaluate Midas with video data from various external data sources and real-world radar data, demonstrating its great advantages over the state-of-the-art approach for both activity recognition and object detection tasks.

Japanese Military “Comfort Women” Knowledge Graph

Materials related to Japanese military “comfort women” in Korea are managed by several institutions. Each digital archive has their own metadata schema and management policies. So far, a standard or a common guideline for describing digital records is not formalized.
 We propose a Japanese military “comfort women” knowledge graph to semantically interlink the digital records from distributed digital archives. To build a Japanese military “comfort women” knowledge graph, digital records and descriptive metadata were collected from existing digital archives. A list of metadata was defined by analyzing commonly used properties and a knowledge model designed by reusing standard vocabularies. Knowledge was constructed by interlinking the collected records, external data sources, and enriching data. The knowledge graph was evaluated using the FAIR data maturity model.

Propensity score stratified MAP prior and posterior inference for incorporating information across multiple potentially heterogeneous data sources

ABSTRACT Incorporation of external information is becoming increasingly common when designing clinical trials. Availability of multiple sources of information has inspired the development of methodologies that account for potential heterogeneity not only between the prospective trial and the pooled external data sources but also between the different external data sources themselves. Our approach proposes an intuitive way of handling such a scenario for the continuous outcomes setting by using propensity score-based stratification and then utilizing robust meta-analytic predictive priors for each stratum to incorporate the prior data to distinguish among different external data sources in each stratum. Through extensive simulations, our approach proves to be more efficient and less biased than the currently available methods. A real case study using clinical trials that study schizophrenia from multiple different sources is also included.