CoChat
Sign Up
Cited by 1Artificial Intelligence in Healthcare and EducationExplainable Artificial Intelligence (XAI)Machine Learning in Healthcare

X-CART: a multimodal real-world data pipeline for explainable AI-driven patient navigation in radiation oncology

Open access
Zink, J.A., Rashid, R., Chinthala, L., Kheirinejad, S., White, B.M., Kheirkhah Rahimabad, P., Hashtarkhani, S., Kumsa, F.A. +4 more·2026·ESMO Real World Data and Digital Oncology
DOI 10.1016/j.esmorw.2026.100771PMC PMC13579299
Open this paper in CoChat — free accountOpen PDF

Abstract

Background: Multimodal explainable artificial intelligence (XAI) is transforming oncology, yet its clinical adoption is constrained by limited interpretability and fragmented use of real-world data (RWD), including electronic medical records, geospatial data, and social determinants of health. We present the XAI CAncer Radiation Therapy (X-CART) platform, a multimodal RWD pipeline designed to enable explainable, clinically actionable precision oncology. Materials and methods: X-CART integrates heterogeneous RWD into a unified, AI-ready architecture for explainable radiotherapy interruption (RTI) risk prediction and patient navigation. A retrospective proof-of-implementation analysis demonstrated the pipeline's predictive modeling and explainability. This framework supports targeted, patient-level intervention strategies. Results: We developed a scalable, interoperable pipeline that synthesizes clinical and community-level data into a secure multimodal framework for RTI risk prediction and patient navigation. The system employs iterative machine learning with feedback-driven retraining and delivers real-time, interpretable outputs through an electronic medical record-embedded dashboard. The architecture is registry-extensible and aligned with interoperability standards to support multi-institutional deployment. As proof of implementation of the Data Science module, a retrospective cohort analysis including 2525 patients identified 622 RTI events (24.6%) and demonstrated modest discrimination with eXtreme Gradient Boosting on an independent hold-out test set (area under the receiver operating characteristic curve = 0.689) with cohort-level SHapley Additive exPlanations interpretability. Conclusion: The framework addresses critical RWD challenges, including data quality, completeness, and traceability. X-CART establishes a generalizable, privacy-preserving foundation for XAI in oncology and provides a practical blueprint for enriching cancer registries with multimodal RWD to support scalable, real-world evidence generation.

AI snapshot

Written from the paper’s full text.

TL;DR

X-CART is a multimodal real-world data pipeline for explainable AI-driven radiotherapy interruption (RTI) risk prediction and patient navigation in radiation oncology. A retrospective proof-of-implementation in 2,525 patients found modest XGBoost discrimination (hold-out AUROC 0.689) with cohort-level SHAP explanations.

Key findings

  • The retrospective cohort included 2,525 radiation oncology patients (August 2020–May 2024), with 622 RTI events (24.6%), defined as two or more missed radiation treatment sessions.
  • On an independent hold-out test set (n = 631), XGBoost achieved AUROC 0.689 (95% CI 0.644–0.734), AUPRC 0.413 (95% CI 0.353–0.501), and Brier score 0.216; at a development-set threshold of 0.455, sensitivity was 0.780, specificity 0.502, and weighted F1 0.597.
  • A class-weighted logistic regression baseline had AUROC 0.678, AUPRC 0.429, Brier score 0.228, sensitivity 0.660, specificity 0.646, and weighted F1 0.672.
  • Cohort-level SHAP attributed predictions mainly to prescribed RT dose (21.1%), dose per fraction (18.8%), and marital status (7.9%), described as model attributions rather than causal effects.

Methods

  • Design: presentation of the X-CART platform architecture plus a retrospective proof-of-implementation of its Data Science module for RTI prediction and explainability.
  • Data: structured clinical, treatment, demographic, behavioral, geospatial, and social determinants of health variables from a single-institution cohort; imaging, radiomics, dosimetry, and clinical text were retained as future extensions and not used in the reported analysis.
  • Analysis: compared a majority-class reference, class-weighted logistic regression, and XGBoost; used training/validation/independent hold-out partitions, stratified 10-fold cross-validation on the development set, AUROC/AUPRC/Brier/sensitivity/specificity/weighted F1, and post hoc cohort-level SHAP.

Limitations

  • Stated: the model was developed in a single-institution cohort with a specific regional case mix and SDoH profile, so feature distributions, calibration, and thresholds may shift across settings; multi-institutional external validation is needed.
  • Stated: the proof of implementation did not include automated treatment delivery, dosimetry, or fraction-level treatment data, and X-CART currently operates within a single-institution EMR environment rather than a population-based cancer registry.
  • Stated: system-level validation is planned but not reported as an empirical outcome; the analysis is predictive rather than causal, and SHAP values are associational. Evident from the text: discrimination was modest and comparable to the logistic regression baseline, and prospective workflow testing had not yet been performed.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1016/j.esmorw.2026.100771

Cites

32 of 32 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

9 of 31 cited sources checked · 2 supported · 1 flagged · accuracy not yet measured

  • Premisecitation

    Clinical adoption of multimodal XAI in oncology is constrained by limited interpretability and fragmented use of RWD.

    “Multimodal explainable artificial intelligence (XAI) is transforming oncology, yet its clinical adoption is constrained by limited interpretability and fragmented use of real-world data (RWD), including electronic medical records, geospatial data, and social determinants of health.”
    Quote in the paperSubstantial claim3 of 3 sources checked
    • cites [1] Papadimitroulas, 2021, Artificial intelligence: deep learning in oncological radiomics and challenges of interpretability and data harmonization, Phys Med source

      Section 1.3 (page unavailable) says interpretability is a challenge that must be addressed to facilitate clinical implementation of deep-learning models in oncological radiomics. Section 1.1 calls for integrating clinical, imaging and other data, but the paper does not establish fragmented use of real-world data—including geospatial data and social determinants of health—as the claimed constraint on adoption. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
    • cites [2] Rao, 2025, The limitations of artificial intelligence in head and neck oncology, Adv Ther source

      The 'Lack of Understanding and Trust' section says opaque AI models impede clinical adoption, and the 'Data-Related Challenges' section describes inconsistent, poorly standardized datasets. The paper discusses XAI as a potential solution, but does not establish that multimodal XAI itself is constrained by fragmented use of real-world data, including electronic medical records, geospatial data, and social determinants of health. No page numbers are available in our copy. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
    • cites [7] Martínez-García, 2022, Data integration challenges for machine learning in precision medicine, Front Med (Lausanne) source

      Section 5 identifies interpretability as an issue for clinical AI/ML and says that integrating electronic health records with social and environmental information remains a challenge. The Introduction also describes the heterogeneity of electronic medical records. These passages support the general barriers, but do not establish them specifically for multimodal XAI in oncology or identify geospatial data as part of fragmented RWD use. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
  • Premisecitation

    Prior studies suggest that upward of 50% of socially vulnerable RT patients may experience treatment interruptions, particularly uninsured patients and Medicaid beneficiaries.

    “prior studies suggest that upward of 50% of socially vulnerable RT patients may experience treatment interruptions, particularly uninsured patients and Medicaid beneficiaries.17, 18, 19”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [17] de Souza, 2017, Determinants of patient-centered financial stress in patients with locally advanced head and neck cancer, J Oncol Pract source
    • cites [18] Lazarev, 2018, Premature discontinuation of curative radiation therapy: insights from head and neck irradiation, Adv Radiat Oncol source

      The Results (Treatment-related characteristics) report RT interruptions in 66% of patients, supporting the ‘upward of 50%’ figure. But these were 58 head-and-neck cancer patients who had already discontinued curative RT prematurely, not socially vulnerable RT patients generally. The Results (Patient demographic and clinical characteristics) list Medicaid enrollment but provide no interruption rate by insurance status and identify no uninsured group. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [19] Thomas, 2017, Interruptions of head and neck radiotherapy across insured and indigent patient populations, J Oncol Pract source
  • Findingcitation

    The Data Architecture Team uses ETL methods to load diverse data sources, including hospital EMR-sourced demographic, behavioral, and clinical data, geospatial data linked to patients' geographic identifiers, and census-tract level SDoH data.

    “The Data Architecture Team uses Extract, Transform, Load (ETL) methods29 to load diverse data sources, including hospital EMR-sourced demographic, behavioral, and clinical data, geospatial data linked to patients’ geographic identifiers, and census-tract level SDoH data.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [29] Wang, 2024, Clinical data warehousing: a scoping review, J Soc Clin Data Manag source

      Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured
  • Premisecitation

    RTI risk is shaped not only by clinical factors but also by social and environmental determinants that are rarely captured in traditional registry data, including transportation barriers, housing instability, and neighborhood-level deprivation.

    “RTI risk is shaped not only by clinical factors but also by social and environmental determinants that are rarely captured in traditional registry data, including transportation barriers, housing instability, and neighborhood-level deprivation.20”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [20] Shaban-Nejad, 2024, Towards an explainable AI platform to study interruptions in cancer radiation therapy, Stud Health Technol Inform source

      The Abstract links social disadvantage with increased radiation therapy interruptions, and the Methods describe combining community-level social determinants with patient and clinical data. The paper does not identify transportation barriers, housing instability, or neighborhood-level deprivation as specific risks, nor does it say these factors are rarely captured in traditional registry data. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
  • Findingown data

    A retrospective proof-of-implementation analysis demonstrated the pipeline's predictive modeling and explainability.

    “A retrospective proof-of-implementation analysis demonstrated the pipeline’s predictive modeling and explainability.”
    Quote in the paperSubstantial claim
  • May combine two claims. 28
  • Classified as probably not a substantial claim. 47

  • May combine two claims.

    Premisecitation

    Many existing XAI tools in Radiation Oncology rely on single-modality data or exclude unstructured data sources such as clinical narratives and imaging.

    “Many existing XAI tools in Radiation Oncology rely on single-modality data24 or exclude unstructured data sources such as clinical narratives and imaging.25,26”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [24] Heising, 2026, Maximizing impact of explainable artificial intelligence in radiotherapy: a critical review, Phys Med Biol source

      Checked before — its earlier verdict is held back until a new check runs.

    • cites [25] Miotto, 2016, Deep patient: an unsupervised representation to predict the future of patients from the electronic health records, Sci Rep source

      The citation is attached to the claim that tools exclude unstructured sources such as clinical narratives and imaging. In Material and Methods > EHR Processing (page unavailable), this paper says it retained free-text clinical notes and extracted concepts from them—the opposite of excluding clinical narratives. The cited passages do not establish a claim about XAI tools in radiation oncology. (LLM, conf=0.88)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
    • cites [26] Rashid, 2025, Machine learning and causal inference-based predictive risk modeling of unplanned radiation treatment interruption [abstract A061], Clin Cancer Res source

  • Classified as probably not a substantial claim.

    Premisecitation

    Multimodal AI platforms incorporate diverse data types to enhance performance of diagnostic and prognostic models.

    “Multimodal AI platforms incorporate diverse data types to enhance the performance of these diagnostic and prognostic models.5”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [5] Lipkova, 2022, Artificial intelligence for multimodal data integration in oncology, Cancer Cell source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    In a related retrospective analysis of 2130 RT patients treated in Eastern Tennessee, nearly 1 in 4 missed two or more scheduled treatments.

    “In a related retrospective analysis of 2130 RT patients treated in Eastern Tennessee, nearly 1 in 4 missed two or more scheduled treatments;”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [21] Kumsa, 2025, Leveraging centralized health system data management and large language model-based data preprocessing to identify predictors for radiation therapy interruption, JCO Clin Cancer Inform source

  • Classified as probably not a substantial claim.

    Findingnone given

    Once validated, an HL7 FHIR interface transmits the model's output back into the hospital EMR system.

    “Once validated, an HL7 FHIR interface transmits the model’s output back into the hospital EMR system.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Findingnone given

    The CHSS-patient relationship is maintained through the early survivorship period for ≥3 months.

    “The CHSS–patient relationship is maintained through the early survivorship period for ≥3 months (Figure 1).”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingnone given

    X-CART is designed to leverage a contextually diverse, AI-ready dataset that prioritizes interpretability and explainability for clinical usability.

    “In contrast, X-CART is designed to leverage a contextually diverse, AI-ready dataset that prioritizes interpretability and explainability for clinical usability.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Findingnone given

    The X-CART framework is being implemented within UTHSC-affiliated oncology clinics, including UTMC.

    “The X-CART framework is being implemented within University of Tennessee Health Science Center (UTHSC)-affiliated oncology clinics, including collaborating clinical sites at the University of Tennessee Medical Center (UTMC).”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Findingnone given

    After automated risk stratification, patients identified to be at elevated risk are referred to a CHSS, who assesses patient-identified social and logistical barriers to treatment, develops individualized navigation plans, and coordinates referrals to community resources.

    “After automated risk stratification, patients identified to be at elevated risk are referred to a CHSS, who assesses patient-identified social and logistical barriers to treatment, develops individualized navigation plans, and coordinates referrals to community resources.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    The User Experience Team is responsible for the design of a CDS dashboard within the EMR that filters and presents real-time data and ML model-generated recommendation alerts.

    “The User Experience Team is responsible for the design of a CDS dashboard within the EMR that filters and presents real-time data and ML model-generated recommendation alerts.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    X-CART is a scalable, interoperable pipeline that integrates heterogeneous clinical and community-level RWD into a secure, unified, AI-ready architecture for explainable RTI risk prediction and patient navigation.

    “We present the XAI CAncer Radiation Therapy (X-CART) platform, a multimodal RWD pipeline designed to enable explainable, clinically actionable precision oncology.”
    • Also stated in: § Results · plain
      “We developed a scalable, interoperable pipeline that synthesizes clinical and community-level data into a secure multimodal framework for RTI risk prediction and patient navigation.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The retrospective cohort's RTI events and features are accurately captured in the EMR and geospatial data.

  • Classified as probably not a substantial claim.

    Findingcitation

    The architecture is designed to return model outputs to the EMR through HL7 FHIR interfaces, where they can be displayed through role-specific dashboards to support point-of-care decision making and patient navigation.

    “The architecture is designed to return model outputs to the EMR through Health Level Seven International (HL7) Fast Healthcare Interoperability Resources (FHIR)28 interfaces, where they can be displayed through role-specific dashboards to support point-of-care decision making and patient navigation.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [28]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Findingnone given

    The system employs iterative machine learning with feedback-driven retraining and delivers real-time, interpretable outputs through an EMR-embedded dashboard.

    “The system employs iterative machine learning with feedback-driven retraining and delivers real-time, interpretable outputs through an electronic medical record-embedded dashboard.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    The X-CART pipeline comprises three operational modules: Data Architecture, Data Science, and User Experience.

    “The X-CART pipeline comprises three operational modules: the Data Architecture Team, which acquires, harmonizes, and deidentifies multimodal data; the Data Science Team, which performs preprocessing, feature engineering, predictive modeling, and explainability analyses; and the User Experience Team, which integrates model outputs into clinician-facing workflows”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Both model and system-level validation metrics are employed to confirm reliability.

    “Both model and system-level validation metrics are employed to confirm reliability.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The finalized analytic cohort was divided into training, validation, and independent hold-out test sets.

    “The finalized analytic cohort was divided into training, validation, and independent hold-out test sets.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Identifying high-risk patient candidates for individualized support interventions is key to maintaining cancer care quality.

    “Therefore, identifying high-risk patient candidates for individualized support interventions is key to maintaining cancer care quality.7”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationreasoning

    The color-coded indicators visually convey the urgency of each AI-generated recommendation and enhance interpretability and explainability of the model outputs, while supporting timely interventions and enabling evaluation of their effectiveness.

    “These color-coded indicators visually convey the urgency of each AI-generated recommendation and enhance interpretability and explainability of the model outputs, while supporting timely interventions and enabling evaluation of their effectiveness.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Identifying patients at risk of RTI through the model and dashboard will enable effective patient navigation and timely interventions.
    • Rests on an assumption: The color-coded risk tiers and alerts will enhance clinician interpretability and decision-making.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductreasoning

    The present proof of implementation did not incorporate automated treatment delivery, dosimetry, or fraction-level treatment data, as these factors are currently outside the scope of the initial pilot build.

    “The present proof of implementation did not incorporate automated treatment delivery, dosimetry, or fraction-level treatment data, as these factors are currently outside the scope of the initial pilot build.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    Although geospatial and SDoH features are conceptually transportable, implementation at other sites will require local geocoding/SDoH linkage, data element mapping, recalibration, and multi-institutional external validation.

    “Although geospatial and SDoH features are conceptually transportable, implementation at other sites will require local geocoding/SDoH linkage, data element mapping, recalibration, and multi-institutional external validation.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    Dashboard elements displayed for each role are designed to support explainability by presenting model outputs in clinically interpretable formats via alerts.

    “Dashboard elements displayed for each role are designed to support explainability by presenting model outputs in clinically interpretable formats via alerts.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The color-coded risk tiers and alerts will enhance clinician interpretability and decision-making.

  • May combine two claims.

    Interpretationown data

    The SHAP analyses provide transparent attribution of risk to specific real-world variables, including SDoH and geospatial factors that are typically absent from registry datasets, thereby demonstrating the added explanatory value of multimodal RWD integration.

    “The SHAP analyses are particularly relevant in the context of RWD-driven oncology research, as they provide transparent attribution of risk to specific real-world variables, including SDoH and geospatial factors that are typically absent from registry datasets, thereby demonstrating the added explanatory value of multimodal RWD integration.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Registry datasets typically lack higher-resolution clinical, geospatial, and SDoH information.

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given

    All study procedures were approved by the UTHSC Institutional Review Board and are conducted in accordance with institutional ethical standards and HIPAA.

    “All study procedures were approved by the UTHSC Institutional Review Board and are conducted in accordance with institutional ethical standards and the Health Insurance Portability and Accountability Act (HIPAA).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductnone given

    The retrospective proof-of-implementation analysis used structured clinical, treatment, demographic, behavioral, geospatial, and SDoH variables, while imaging, radiomics, and dosimetry remain as future pipeline extensions.

    “The retrospective proof-of-implementation analysis used structured clinical, treatment, demographic, behavioral, geospatial, and SDoH variables, while imaging, radiomics, and dosimetry remain as future pipeline extensions.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    These SHAP values represent model attributions, not causal effects, illustrating the explanatory value of multimodal RWD integration.

    “These values represent model attributions, not causal effects, illustrating the explanatory value of multimodal RWD integration.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Doing so effectively requires data pipelines capable of integrating and harmonizing diverse RWD streams in a privacy-preserving, AI-ready format.

    “Doing so effectively requires data pipelines capable of integrating and harmonizing these diverse RWD streams in a privacy-preserving, AI-ready format.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    XAI frameworks such as LIME and SHAP provide interpretable visualizations of model predictions.

    “To address this, explainable artificial intelligence (XAI) frameworks such as Local Interpretable Model Agnostic Explanations and SHapley Additive exPlanations (SHAP) have emerged, providing interpretable visualizations of model predictions.3”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [3] Ladbury, 2022, Utilization of model-agnostic explainable artificial intelligence frameworks in oncology: a narrative review, Transl Cancer Res source

      Checked before — its earlier verdict is held back until a new check runs.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    A data compliance checkpoint identifies noncompliant or low-quality records for manual review or automated reprocessing before model development.

    “A data compliance checkpoint identifies noncompliant or low-quality records for manual review or automated reprocessing before model development.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    AI-generated predictions and explainability outputs can be encoded as FHIR Observation resources linked to relevant clinical records, facilitating interoperability, registry-compatible data exchange, and integration into clinical workflows.

    “AI-generated predictions and explainability outputs can be encoded as FHIR Observation resources linked to relevant clinical records, facilitating interoperability, registry-compatible data exchange, and integration into clinical workflows.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    Role-based access control and audit logging mechanisms restrict and monitor access to sensitive health data.

    “Role-based access control and audit logging mechanisms restrict and monitor access to sensitive health data.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductown data

    Calibration was assessed using the Brier scores on the validation cohort to evaluate agreement between predicted probabilities and observed RTI frequencies.

    “Calibration was assessed using the Brier scores on the validation cohort to evaluate agreement between predicted probabilities and observed RTI frequencies.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Feature distributions, calibration, and risk thresholds may shift across settings.

    “Feature distributions, calibration, and risk thresholds may therefore shift across settings.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Our group's prior work in West Tennessee found similar treatment-access barriers by neighborhood, race, and insurance status in Memphis.

    “These findings echo our group’s prior work in West Tennessee, which found similar treatment-access barriers by neighborhood, race, and insurance status in Memphis22”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [22] Wakefield, 2020, Location as destiny: identifying geospatial disparities in radiation treatment interruption by neighborhood, race, and insurance, Int J Radiat Oncol Biol Phys source

      Checked before — its earlier verdict is held back until a new check runs.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    The framework addresses critical RWD challenges, including data quality, completeness, and traceability.

    “The framework addresses critical RWD challenges, including data quality, completeness, and traceability.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    RTIs, especially total interruptions of >5 days, lead to accelerated tumor cell repopulation and radioresistance and poorer clinical outcomes, particularly for fast-growing carcinomas such as head and neck cancers.

    “RTIs, especially total interruptions of >5 days, lead to accelerated tumor cell repopulation and radioresistance9 and poorer clinical outcomes,10, 11, 12, 13, 14 particularly for fast-growing carcinomas such as head and neck cancers.15”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [9] Tarnawski, 2002, How fast is repopulation of tumor cells during the treatment gap?, Int J Radiat Oncol Biol Phys source
    • cites [10] Pajak, 1991, Elapsed treatment days--a critical item for radiotherapy quality control review in head and neck trials: RTOG report, Int J Radiat Oncol Biol Phys source
    • cites [11] Robertson, 1998, Similar decreases in local tumor control are calculated for treatment protraction and for interruptions in the radiotherapy of carcinoma of the larynx in four centers, Int J Radiat Oncol Biol Phys source

  • Classified as probably not a substantial claim.

    Premisecitation

    Integration of varied data sources, including imaging, dosimetry, and EMR, has enabled creation of comprehensive predictive models in Radiation Oncology.

    “The integration of varied data sources, including imaging, dosimetry, and electronic medical record (EMR), has enabled the creation of comprehensive predictive models in Radiation Oncology.6”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [6] Bibault, 2016, Big Data and machine learning in radiation oncology: state of the art and future prospects, Cancer Lett source

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationreasoning

    The use of HL7 FHIR as the interoperability standard ensures that model outputs can, in principle, be federated to external systems, including cancer registry platforms, thereby supporting broader RWD integration efforts.

    “The use of HL7 FHIR as the interoperability standard ensures that model outputs can, in principle, be federated to external systems, including cancer registry platforms, thereby supporting broader RWD integration efforts.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The threshold used for hold-out reporting was selected from development set out-of-fold predictions to prioritize detection of the interruption class.

    “The threshold used for hold-out reporting was selected from development set out-of-fold predictions to prioritize detection of the interruption class.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    Challenges remain in data harmonization and generalizability of AI models in Radiation Oncology, such as small dataset size and varying feature extraction sensitivity.

    “Although promising, challenges remain in data harmonization and the generalizability of AI models in this area, such as small dataset size and varying feature extraction sensitivity.1”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [1] Papadimitroulas, 2021, Artificial intelligence: deep learning in oncological radiomics and challenges of interpretability and data harmonization, Phys Med source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    In prior work, our group demonstrated an emerging AI predictive model platform that has the potential to accurately triage RTI risk and direct personalized patient support to reduce barriers to receiving RT.

    “Furthermore, in prior work, our group demonstrated an emerging AI predictive model platform that has the potential to accurately triage radiotherapy interruption (RTI) risk and direct personalized patient support to reduce barriers to receiving radiotherapy (RT).8”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [8] Rashid, 2024 source

  • May combine two claims.

    Interpretationown data

    The results demonstrate the feasibility of the predictive modeling component of the pipeline, including benchmarking conventional and nonlinear models and supporting explainability through the XGBoost-SHAP component, rather than establishing a deployment-ready clinical prediction model.

    “Together, these results demonstrate the feasibility of the predictive modeling component of the pipeline, including benchmarking conventional and nonlinear models and supporting explainability through the XGBoost-SHAP component, rather than establishing a deployment-ready clinical prediction model.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    The development set, composed of the training and validation partitions, was also evaluated using stratified 10-fold cross-validation, while the independent hold-out test set was reserved for final model evaluation.

    “The development set, composed of the training and validation partitions, was also evaluated using stratified 10-fold cross-validation, while the independent hold-out test set was reserved for final model evaluation.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    In that analysis, Medicaid coverage, marital instability, and high neighborhood social vulnerability index scores predicted interruption risk.

    “Medicaid coverage, marital instability, and high neighborhood social vulnerability index scores predicted interruption risk”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [21] Kumsa, 2025, Leveraging centralized health system data management and large language model-based data preprocessing to identify predictors for radiation therapy interruption, JCO Clin Cancer Inform source

      Checked before — its earlier verdict is held back until a new check runs.

  • May combine two claims.

    Premisecitation

    The complexity of AI models can result in a 'black box' problem, in which lack of explanatory transparency can limit provider confidence and clinical adoption.

    “However, the complexity of these models can result in a ‘black box’ problem, in which a lack of explanatory transparency driving recommendations can limit provider confidence in results and clinical adoption.2”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [2] Rao, 2025, The limitations of artificial intelligence in head and neck oncology, Adv Ther source

      Checked before — its earlier verdict is held back until a new check runs.

  • Classified as probably not a substantial claim.

    Interpretationreasoning

    The X-CART framework supports targeted, patient-level intervention strategies.

    “This framework supports targeted, patient-level intervention strategies.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Geospatial mapping tied elevated risk to specific rural ZIP codes.

    “geospatial mapping tied elevated risk to specific rural ZIP codes.21”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [21] Kumsa, 2025, Leveraging centralized health system data management and large language model-based data preprocessing to identify predictors for radiation therapy interruption, JCO Clin Cancer Inform source

      Checked before — its earlier verdict is held back until a new check runs.

  • May combine two claims.

    Interpretationown data

    XGBoost achieved modest discrimination on an independent hold-out test set, with performance comparable to baseline, while enabling cohort-level SHAP attributions spanning treatment intensity, demographic, clinical, and SDoH domains.

    “Our pilot work confirms that XGBoost achieved modest discrimination on an independent hold-out test set, with performance comparable to baseline, while enabling cohort-level SHAP attributions spanning treatment intensity, demographic, clinical, and SDoH domains.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    A recent review of XAI applications in oncology emphasizes the need for interpretability and trustworthiness in AI-mediated tools in health care.

    “A recent review of XAI applications in oncology emphasizes the need for interpretability and trustworthiness in AI-mediated tools as deployed in health care.4”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [4] Ghasemi, 2024, Explainable artificial intelligence in breast cancer detection and risk prediction: a systematic scoping review, Cancer Innov source

      Checked before — its earlier verdict is held back until a new check runs.

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationown data

    These findings support the feasibility of the pipeline while underscoring need for prospective workflow testing and multi-institutional validation before deployment.

    “These findings support the feasibility of the pipeline while underscoring need for prospective workflow testing and multi-institutional validation before deployment.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Identifying patients at risk of RTI through the model and dashboard will enable effective patient navigation and timely interventions.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Interpretationreasoning

    The architecture is designed to support modular security configurations and federated learning extensions, enabling collaborative model training across firewalled institutions without centralizing raw patient data.

    “In anticipation of scaling to multi-institutional deployments, the architecture is designed to support modular security configurations and federated learning extensions, enabling collaborative model training across firewalled institutions without centralizing raw patient data.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductnone given

    This study is single-institution and retrospective; the independent hold-out test set approximates out-of-sample performance.

    “This study is single-institution and retrospective; the independent hold-out test set approximates out-of-sample performance.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Interpretationnone given

    The architecture incorporates explainability as a core design principle, ensuring that AI-generated outputs are understandable, interpretable, and actionable for oncologists, care coordinators, and CHSSs.

    “The architecture incorporates explainability as a core design principle, ensuring that AI-generated outputs are understandable, interpretable, and actionable for oncologists, care coordinators, and community health support specialists (CHSSs).”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductreasoning

    In this report, the empirical analysis is predictive rather than causal, with predictions explained by SHAP as associational feature attributions rather than causal effects.

    “In this report, the empirical analysis is predictive rather than causal, with predictions explained by SHAP as associational feature attributions rather than causal effects.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Class imbalance was addressed using class weighting.

    “Class imbalance was addressed using class weighting.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationnone given

    The architecture is registry-extensible and aligned with interoperability standards to support multi-institutional deployment.

    “The architecture is registry-extensible and aligned with interoperability standards to support multi-institutional deployment.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The X-CART architecture's design principles are sufficient to support multi-institutional deployment and registry enrichment.

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given

    Records missing the RTI outcome were excluded, patients with >50% missing data were removed, minor categorical missingness was labeled 'unknown,' outliers were capped at the 1st and 99th percentiles, and highly skewed variables were log-transformed.

    “Records missing the RTI outcome were excluded, patients with >50% missing data were removed, minor categorical missingness was labeled ‘unknown,’ outliers were capped at the 1st and 99th percentiles, and highly skewed variables were log-transformed.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    Deidentified data used for model training is stored in a secure, access-controlled research enclave, isolated from EMR environments.

    “Deidentified data used for model training is stored in a secure, access-controlled research enclave, isolated from EMR environments.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    Preventing interruptions can be challenging given that RT is time-consuming (up to 6-7 weeks), toxic, and generates high out-of-pocket patient expense.

    “Preventing interruptions can be challenging given that RT is time-consuming (up to 6-7 weeks), toxic, and generates high out-of-pocket patient expense16;”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [16] Davidoff, 2013, Out-of-pocket health care expenditure burden for Medicare beneficiaries with cancer, Cancer source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Before preprocessing, all patient data undergoes deidentification and encryption to support HIPAA compliance.

    “Before preprocessing, all patient data undergoes deidentification and encryption to support HIPAA compliance.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Architectures that can bridge registry-level population data with institution-level multimodal clinical data are increasingly needed.

    “Architectures that can bridge registry-level population data with institution-level multimodal clinical data are therefore increasingly needed.”
    Quote in the paperNot a substantial claim
59 of 59

  • The paper relies on this without support.

    The retrospective cohort's RTI events and features are accurately captured in the EMR and geospatial data.

    “Input data include individual-level EMR data such as clinical text, SDoH variables, and population-level geospatial data, which are RWD sources often siloed and underutilized in registry-based oncology research.”

  • The paper relies on this without support.

    The XGBoost model's performance on the hold-out test set is a valid estimate of its performance in the implementation setting.

    “The finalized analytic cohort was divided into training, validation, and independent hold-out test sets.”

  • The paper relies on this without support.

    SHAP values provide clinically meaningful explanations of RTI risk.

    “In this report, the empirical analysis is predictive rather than causal, with predictions explained by SHAP as associational feature attributions rather than causal effects.”

  • The paper relies on this without support.

    The X-CART architecture's design principles are sufficient to support multi-institutional deployment and registry enrichment.

    “The architecture is registry-extensible and aligned with interoperability standards to support multi-institutional deployment.”

  • The paper relies on this without support.

    Registry datasets typically lack higher-resolution clinical, geospatial, and SDoH information.

    “including SDoH and geospatial factors that are typically absent from registry datasets”

  • The paper relies on this without support.

    Identifying patients at risk of RTI through the model and dashboard will enable effective patient navigation and timely interventions.

    “supporting timely interventions and enabling evaluation of their effectiveness”

  • The paper relies on this without support.

    The color-coded risk tiers and alerts will enhance clinician interpretability and decision-making.

    “enhance interpretability and explainability of the model outputs”

  • The paper relies on this without support.

    The binary RTI outcome, defined as two or more missed radiation treatment sessions, is a clinically meaningful and valid measure of radiotherapy interruption.

    “defined as two or more missed radiation treatment sessions”
8 of 8

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. 2024, About PLACES: local data for better health, CDC PLACES

Cited by

Sorted by citations

Related

Authors
Zink, J. A., Rashid, R., Chinthala, L., Kheirinejad, S., White, B. M., Kheirkhah Rahimabad, P., Hashtarkhani, S., Kumsa, F. A., Brett, C. L., Davis, R. L., Schwartz, D. L., & Shaban-Nejad, A.
Affiliations
  • Oak Ridge National Laboratory (US)
  • University of Tennessee at Knoxville (US)
  • University of Tennessee Health Science Center (US)
Corresponding author
Schwartz, D.L.
ORCID
Venue
ESMO Real World Data and Digital Oncology
Publisher
Elsevier BV
Published
2026-09
Volume
13
Pages
100771
ISSN
2949-8201 (ISSN-L)
DOI
10.1016/j.esmorw.2026.100771
PMC
PMC13579299
OpenAlex
W7211885082
Cited by
1
References
27
Citations by year
2026: 1
FWCI
1.95
Type
journal article
Language
en
Open access
Licence
Locations
  • ESMO Real World Data and Digital Oncology (publishedVersion, cc-by-nc-nd)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
Funders
Tennessee Department of Health
Topics
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
Keywords
explainable AI, radiotherapy interruption, patient navigation, radiation oncology, precision oncology, multimodal data integration, electronic health records, social determinants of health, geospatial data, cancer registry, SHAP, XGBoost, privacy-preserving data integration, interoperability standards, real-world evidence

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-11-06.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-11-06.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review