Medical ML: Open Questions for Future Research — A Medical AI Research Agenda for Ukrainian Healthcare

Posted on by

Open Questions for Future Research: A Medical AI Research Agenda for Ukrainian Healthcare

By Oleh Ivchenko, PhD Candidate | Odessa National Polytechnic University | Stabilarity Hub | February 11, 2026

Abstract

After twelve weeks examining machine learning applications in medical imaging diagnosis, significant knowledge gaps remain that demand systematic investigation. This concluding article synthesizes open research questions emerging from our comprehensive review, organized across seven priority domains: generalization and distribution shift, algorithmic fairness and bias mitigation, human-AI collaboration optimization, regulatory science advancement, healthcare system integration, emerging model architectures, and low-resource context adaptation. We present a structured research agenda positioning Ukrainian healthcare institutions as contributors to global medical AI advancement while addressing locally relevant challenges. Each research domain includes specific questions amenable to empirical investigation, methodological considerations, and potential collaboration opportunities. The agenda particularly emphasizes questions relevant to middle-income healthcare systems navigating AI integration with constrained resources—a context underrepresented in current literature. This article serves as both a conclusion to our research series and a roadmap for continued scientific inquiry into the transformative potential of AI-augmented medical diagnosis.

35
articles published in this research series
7
priority research domains identified
42
specific research questions proposed
222+
commercial AI products in medical imaging (Oct 2024)

1. Introduction

This article concludes a systematic twelve-week research program examining machine learning applications in medical imaging diagnosis for Ukrainian healthcare. Across 34 preceding articles, we analyzed global best practices, documented failed implementations, evaluated technical architectures, and developed practical frameworks for AI integration into clinical workflows. Yet every answer generated new questions—the hallmark of productive scientific inquiry.

Medical AI has evolved from laboratory curiosity to clinical reality, with 222+ commercial products available and 1,200+ FDA authorizations as of late 2024. However, this proliferation masks fundamental uncertainties: real-world validation remains sparse, deployment outside high-resource settings is poorly characterized, and the optimal modes of human-AI collaboration are still debated. A 2025 Lancet rapid scoping review concluded that “the potential benefits of AI are evident, but there is a paucity of evidence in real-world settings.”

For Ukrainian healthcare institutions—and similarly positioned systems in middle-income countries—these uncertainties compound with resource constraints, infrastructure limitations, and the need to adapt technologies developed primarily for Western contexts. This creates both challenges and opportunities: challenges in implementing AI responsibly with limited evidence, opportunities in contributing research that addresses globally underrepresented contexts.

1.1 Purpose and Scope

This research agenda aims to:

  • Synthesize open questions emerging from our comprehensive literature review
  • Prioritize research directions with highest potential for advancing medical AI implementation
  • Identify questions particularly relevant to Ukrainian and middle-income healthcare contexts
  • Propose methodological approaches and collaboration opportunities for each domain
  • Position ScanLab and affiliated institutions as contributors to global medical AI research

2. Research Domain Overview

flowchart TB
    subgraph Core["Core Technical Domains"]
        A[Generalization &
Distribution Shift]
        B[Fairness &
Bias Mitigation]
        C[Emerging
Architectures]
    end
    
    subgraph Applied["Applied Research Domains"]
        D[Human-AI
Collaboration]
        E[Healthcare
Integration]
        F[Regulatory
Science]
    end
    
    subgraph Context["Context-Specific"]
        G[Low-Resource
Adaptation]
    end
    
    A --> D
    B --> D
    C --> D
    D --> E
    E --> F
    F --> G
    G --> A
    
    style G fill:#ff9800,color:#000
Table 1: Research Domain Summary
Domain Key Challenge Questions Priority
Generalization & Distribution Shift Models fail when deployment differs from training 8 Critical
Fairness & Bias Disparate performance across populations 6 Critical
Human-AI Collaboration Optimizing combined human-machine performance 7 High
Regulatory Science Appropriate oversight for adaptive systems 5 High
Healthcare Integration Real-world implementation challenges 6 High
Emerging Architectures Foundation models, multimodal systems 5 Medium
Low-Resource Adaptation Deployment in resource-constrained settings 5 Critical (for Ukraine)

3. Domain 1: Generalization and Distribution Shift

The fundamental challenge of medical AI deployment is generalization: models trained on specific datasets frequently fail when applied to populations, equipment, or protocols that differ from training conditions. This “distribution shift” problem undermines the validity of reported performance metrics and creates patient safety risks.

🔬 Research Priority Statement

Developing robust methods for detecting, measuring, and adapting to distribution shift is the single most important technical challenge for medical AI deployment.

3.1 Open Research Questions

Table 2: Generalization Research Questions
# Question Current State Approach
G1 What metrics best predict when a trained model will fail on new data? Uncertainty quantification emerging but not validated Prospective deployment studies with stratified analysis
G2 How much local data is required for effective domain adaptation? Varies widely; no systematic characterization Multi-site transfer learning experiments
G3 Can we detect distribution shift in real-time during clinical deployment? Prototype monitoring systems exist; clinical validation lacking Prospective monitoring with ground truth collection
G4 What imaging acquisition parameters most strongly influence model generalization? Manufacturer, resolution known factors; comprehensive mapping absent Controlled variation studies across equipment types
G5 How do federated learning models generalize compared to centrally trained models? Theoretical advantages; real-world comparisons sparse Matched federated vs. centralized training experiments
G6 What is the optimal frequency for model retraining or recalibration? Ad hoc approaches; no evidence-based guidelines Longitudinal performance monitoring with retraining triggers
G7 Can synthetic data augmentation improve generalization to unseen domains? Mixed results; context-dependent effectiveness Systematic augmentation strategy comparison
G8 How do models trained on Western populations perform on Ukrainian patient demographics? Unknown; no published Ukrainian validation studies External validation on Ukrainian datasets

3.2 Ukrainian Research Opportunities

Question G8 represents a high-priority opportunity for Ukrainian researchers. The absence of published validation studies on Eastern European populations creates a gap that local institutions can uniquely address. Proposed approach:

  1. Partner with ScanLab implementation sites to collect prospective data
  2. Apply leading FDA-cleared algorithms to Ukrainian imaging datasets
  3. Report stratified performance by demographic and clinical characteristics
  4. Characterize distribution shift from training populations (typically US/EU)
  5. Develop domain adaptation protocols if performance degradation observed

4. Domain 2: Algorithmic Fairness and Bias Mitigation

AI systems can perpetuate and amplify existing healthcare disparities. Representation bias in training data, surrogate endpoint selection, and differential access to AI-enabled care create risks of inequitable outcomes. The 2025 NCBI Watch List noted that “existing health care data can lead to AI systems perpetuating biases and exacerbating disparities.”

flowchart LR
    subgraph Sources["Bias Sources"]
        A[Historical Data
Reflects past disparities]
        B[Representation
Training population limits]
        C[Measurement
Proxy variable bias]
        D[Aggregation
Hidden subgroup effects]
    end
    
    subgraph Manifestation["Manifestation"]
        E[Differential
Performance]
        F[Systematically
Worse Outcomes]
    end
    
    subgraph Impact["Patient Impact"]
        G[Delayed
Diagnosis]
        H[Inappropriate
Treatment]
        I[Widened
Disparities]
    end
    
    A --> E
    B --> E
    C --> F
    D --> F
    E --> G
    F --> H
    G --> I
    H --> I

4.1 Open Research Questions

Table 3: Fairness Research Questions
# Question Significance
F1 What fairness metrics are most clinically meaningful for medical AI? Statistical fairness may conflict with clinical utility
F2 How can we audit deployed AI systems for bias without demographic labels in clinical data? Many EHR systems lack complete demographic information
F3 What minimum dataset diversity is required for equitable performance across populations? Practical guidance for dataset curation lacking
F4 Can post-hoc fairness interventions (threshold adjustment, recalibration) adequately address training data bias? Trade-offs between approaches poorly characterized
F5 How do fairness considerations interact with regulatory approval processes? FDA guidance emerging but implementation unclear
F6 What are the unique bias risks for medical AI deployed in Eastern European contexts? Most bias research focuses on US racial categories

📊 Ukrainian Context

Question F6 highlights a significant research gap. Western bias frameworks focus on US-centric categories (Black/White/Hispanic), which may not capture relevant disparities in Ukrainian healthcare. Alternative dimensions requiring investigation include:

  • Urban-rural disparities: Differential access to modern imaging equipment
  • Age-related demographics: Older population with different disease presentations
  • Socioeconomic factors: Access to private vs. public healthcare systems
  • Historical exposures: Chornobyl-related health patterns, tuberculosis prevalence

5. Domain 3: Human-AI Collaboration Optimization

Medical AI systems are not autonomous—they augment human decision-making. Yet optimal collaboration modes remain poorly understood. Over-reliance (automation bias) and under-reliance both degrade combined performance. The fundamental question: how do we achieve human-AI teams that outperform either alone?

flowchart TD
    subgraph Input["Clinical Input"]
        A[Medical Image]
        B[Clinical Context]
    end
    
    subgraph Processing["Parallel Processing"]
        C[AI Analysis]
        D[Human Analysis]
    end
    
    subgraph Integration["Integration Challenge"]
        E{How to Combine?}
    end
    
    subgraph Outcomes["Performance Outcomes"]
        F[Combined > Either Alone?]
        G[Trust Calibration?]
        H[Cognitive Load?]
    end
    
    A --> C
    A --> D
    B --> D
    C --> E
    D --> E
    E --> F
    E --> G
    E --> H

5.1 Open Research Questions

Table 4: Human-AI Collaboration Research Questions
# Question Research Approach
H1 Under what conditions does AI assistance improve versus degrade physician diagnostic accuracy? Randomized trials with stratification by case difficulty and physician experience
H2 What interface designs promote appropriate reliance (neither over- nor under-trust)? Usability studies with trust calibration metrics
H3 How should AI confidence/uncertainty be communicated to support clinical decision-making? A/B testing of uncertainty visualization approaches
H4 Does AI-first versus AI-second workflow sequencing affect diagnostic outcomes? Crossover studies comparing workflow variants
H5 What training interventions most effectively calibrate physician trust in AI? Educational intervention trials with longitudinal follow-up
H6 How do time pressures affect physician interaction with AI recommendations? Observational studies with workload tracking
H7 What explainability methods most effectively support physician override decisions? Comparative studies of explanation types (attention maps, concept-based, etc.)

6. Domain 4: Regulatory Science Advancement

Current regulatory frameworks were designed for static medical devices, not adaptive AI systems that may learn continuously from deployment data. The FDA’s 2021 AI/ML-Based Software as Medical Device Action Plan acknowledged this gap but implementation details remain uncertain.

⚖️ Regulatory Challenge

How do we balance innovation-enabling flexibility with safety-ensuring oversight for AI systems whose behavior may change post-approval?

6.1 Open Research Questions

Table 5: Regulatory Science Research Questions
# Question Stakeholders
R1 What post-market surveillance methodologies adequately detect performance degradation? Regulators, healthcare systems, AI vendors
R2 How should predetermined change control plans specify “locked” boundaries for continuous learning? FDA/notified bodies, developers
R3 What evidence standards should apply to AI systems validated on synthetic or augmented data? Regulators, researchers
R4 How can regulatory harmonization be achieved across FDA, EU MDR, and emerging frameworks (e.g., Ukrainian MHSU)? International regulators, global vendors
R5 What liability frameworks appropriately allocate responsibility for AI-assisted diagnostic errors? Legal systems, insurers, clinicians

7. Domain 5: Healthcare System Integration

Even technically excellent AI systems may fail in deployment due to workflow integration challenges. The 81% of hospitals without AI utilization—despite available tools—suggests implementation science questions are as important as technical ones.

7.1 Open Research Questions

Table 6: Integration Research Questions
# Question Measurement Approach
I1 What organizational factors predict successful AI adoption in imaging departments? Multi-site implementation studies with organizational surveys
I2 How does AI integration affect radiologist workload in real-world settings (efficiency gains vs. alert fatigue)? Time-motion studies, cognitive load assessment
I3 What PACS integration architectures optimize AI utilization while minimizing workflow disruption? Comparative implementation studies across architecture patterns
I4 How do patients respond to disclosure of AI involvement in their diagnosis? Patient surveys, communication effectiveness studies
I5 What change management approaches most effectively reduce physician resistance to AI adoption? Intervention trials with adoption and utilization metrics
I6 How should AI integration be sequenced across diagnostic modalities for maximum impact? Phased implementation studies with outcome tracking

8. Domain 6: Emerging Model Architectures

Foundation models and multimodal architectures represent the next frontier in medical AI. These systems—trained on massive datasets and capable of cross-task generalization—may fundamentally change how AI is deployed in healthcare. A 2025 review noted the “shift toward generalizable foundational models” as a defining trend.

flowchart LR
    subgraph Past["2012-2020"]
        A[Task-Specific CNNs]
    end
    
    subgraph Present["2020-2025"]
        B[Vision Transformers]
        C[Transfer Learning]
    end
    
    subgraph Emerging["2025+"]
        D[Foundation Models]
        E[Multimodal Integration]
        F[Generative AI]
    end
    
    subgraph Questions["Open Questions"]
        G[Clinical Utility?]
        H[Safety Guarantees?]
        I[Regulation?]
    end
    
    A --> B
    B --> D
    C --> E
    D --> G
    E --> H
    F --> I

8.1 Open Research Questions

Table 7: Emerging Architecture Research Questions
# Question Technical Challenge
E1 Do medical foundation models outperform task-specific models for clinical diagnostic tasks? Fair comparison methodology with matched computational resources
E2 How should multimodal AI systems integrating imaging with clinical data be validated? Dataset requirements, performance attribution across modalities
E3 What are the hallucination risks of generative AI in medical imaging contexts? Detecting plausible but incorrect generated content
E4 Can LLM-based reasoning enhance clinical interpretation of AI imaging findings? Integration architecture, reliability assessment
E5 What computational infrastructure is minimally required for foundation model deployment? Resource profiling for edge deployment in resource-limited settings

9. Domain 7: Low-Resource Context Adaptation

Most medical AI research and development occurs in high-income settings. This creates systematic gaps for middle- and low-income healthcare systems that must adapt rather than adopt AI technologies. Ukraine’s position as a middle-income country with strong technical capacity creates an opportunity to lead research addressing this underserved area.

🌍 Global Health Relevance

Research addressing AI deployment in resource-constrained settings has outsized impact potential—the majority of the world’s population receives healthcare in contexts more similar to Ukraine than to the US/EU systems where most AI development occurs.

9.1 Open Research Questions

Table 8: Low-Resource Adaptation Research Questions
# Question Ukrainian Relevance
L1 What is the minimum dataset size for effective local fine-tuning of pre-trained medical AI models? Critical for ScanLab adaptation with limited local data
L2 How do infrastructure limitations (connectivity, computational resources) affect AI deployment strategies? Varies significantly across Ukrainian regions
L3 Can AI systems designed for modern equipment effectively process images from older/legacy scanners? Many Ukrainian facilities use older imaging equipment
L4 What cost-effectiveness thresholds justify AI investment in middle-income healthcare systems? Essential for Ukrainian health economics analysis
L5 How should AI curricula be adapted for healthcare workforces with limited prior AI exposure? Training curriculum design for Ukrainian physicians

10. Research Methodology Recommendations

10.1 Study Design Priorities

Based on our review, several methodological approaches are underutilized in current medical AI research:

  • Prospective real-world validation: Most published studies use retrospective data; prospective deployment studies urgently needed
  • Randomized implementation trials: Compare AI-assisted versus standard care with appropriate blinding
  • Multi-site external validation: Test generalization across diverse healthcare settings
  • Longitudinal outcome tracking: Assess long-term patient outcomes, not just immediate diagnostic accuracy
  • Health economic evaluation: Integrate cost-effectiveness analysis into AI evaluation studies

10.2 Reporting Standards

Adoption of standardized reporting frameworks will strengthen the evidence base:

Table 9: Recommended Reporting Standards
Framework Application Status
CLAIM 2024 AI imaging research reporting Updated, widely adopted
CONSORT-AI AI clinical trials Published, adoption growing
TRIPOD+AI Prediction model development Published 2024
SPIRIT-AI AI intervention trial protocols Published, recommended
DECIDE-AI Early clinical evaluation Emerging framework

11. Ukrainian Research Ecosystem Positioning

11.1 Institutional Collaboration Framework

flowchart TB
    subgraph Academic["Academic Partners"]
        A[ONPU Economics
Cybernetics]
        B[Medical
Universities]
        C[Computer Science
Departments]
    end
    
    subgraph Clinical["Clinical Partners"]
        D[Regional
Hospitals]
        E[Diagnostic
Centers]
        F[ScanLab Pilot
Sites]
    end
    
    subgraph Industry["Industry Partners"]
        G[AI Vendors]
        H[PACS Providers]
        I[Stabilarity Hub]
    end
    
    subgraph International["International Network"]
        J[EU Research
Consortia]
        K[WHO Digital
Health]
        L[RSNA/ESR
Research]
    end
    
    A <--> D
    B <--> E
    C <--> I
    F <--> I
    I <--> J
    A <--> K
    B <--> L

11.2 Priority Research Projects

Based on the open questions and Ukrainian context, we propose five priority research projects:

  1. Ukrainian External Validation Study: Systematic evaluation of FDA-cleared AI systems on Ukrainian patient populations (Questions G8, F6)
  2. ScanLab Implementation Science: Prospective study of AI integration workflow effects at pilot sites (Questions I1, I2, H1)
  3. Legacy Equipment Adaptation: Characterizing AI performance on older imaging equipment common in Ukrainian facilities (Question L3)
  4. Ukrainian AI Curriculum Effectiveness: Randomized evaluation of training program impact on appropriate AI utilization (Questions H5, L5)
  5. Cost-Effectiveness in Middle-Income Context: Health economic analysis of AI deployment in Ukrainian healthcare system (Question L4)

12. Conclusion: A Research Manifesto

After twelve weeks examining machine learning in medical imaging, we are left with more questions than answers—a sign of productive scientific inquiry. The 42 open questions presented across seven domains represent not a failure of current research but an honest assessment of the maturation stage of medical AI.

Key conclusions from our research agenda:

  • Generalization remains the fundamental challenge. Until we can reliably predict when AI systems will fail, deployment carries irreducible uncertainty.
  • Human-AI collaboration is under-studied. We have invested heavily in algorithm development but insufficiently in understanding how humans actually interact with AI recommendations.
  • Regulatory frameworks must evolve. Current oversight mechanisms are inadequate for adaptive, continuously learning systems.
  • Middle-income contexts are underrepresented. Most research reflects high-resource settings; Ukraine and similar countries can contribute unique perspectives.
  • Implementation science matters. Technical performance means nothing if AI systems are not adopted and used appropriately in clinical practice.

📢 Call to Action

We invite Ukrainian researchers, clinicians, and healthcare administrators to join this research agenda. The questions outlined here require collaborative effort across institutions and disciplines. Through rigorous investigation of these open problems, Ukraine can position itself as a contributor to—not merely a consumer of—global medical AI advancement.

The 35 articles in this series have laid groundwork; the research agenda presented here charts the path forward. Medical AI’s transformative potential for Ukrainian healthcare will only be realized through sustained, systematic inquiry into the questions that remain unanswered.

References

  1. Lancet EClinicalMedicine. Artificial intelligence for diagnostics in radiology practice: a rapid systematic scoping review. eClinicalMedicine. 2025. DOI: 10.1016/S2589-5370(25)00160-9
  2. NCBI. 2025 Watch List: Artificial Intelligence in Health Care. NCBI Bookshelf. 2025. Available at: ncbi.nlm.nih.gov/books/NBK613808
  3. Kowalczuk M, et al. Artificial Intelligence-Empowered Radiology—Current Status and Critical Review. Diagnostics. 2025;15:282. DOI: 10.3390/diagnostics15030282
  4. Raji ID, et al. AI-driven healthcare: Fairness in AI healthcare: A survey. PLOS Digital Health. 2025. DOI: 10.1371/journal.pdig.0000864
  5. Chen IY, et al. Algorithm fairness in artificial intelligence for medicine and healthcare. Nature Medicine. 2023;29:2933-2942. DOI: 10.1038/s41591-023-02628-3
  6. Rajpurkar P, et al. The future of multimodal artificial intelligence models for integrating imaging and clinical metadata. Diagnostic Interventional Radiology. 2025. DOI: 10.4274/dir.2024.242631
  7. Haug CJ, Drazen JM. Artificial Intelligence and Machine Learning in Clinical Medicine, 2023. NEJM. 2023;388:1201-1208. DOI: 10.1056/NEJMra2302038
  8. Mongan J, et al. Checklist for Artificial Intelligence in Medical Imaging (CLAIM): 2024 Update. Radiology: AI. 2024;6:e240300. DOI: 10.1148/ryai.240300
  9. FDA. Artificial Intelligence and Machine Learning (AI/ML)-Based Software as a Medical Device Action Plan. 2021. Available at: fda.gov/medical-devices
  10. WHO. Ethics and Governance of Artificial Intelligence for Health. WHO Guidance. 2021. ISBN: 978-92-4-003575-9
  11. Topol EJ. High-Performance Medicine: The Convergence of Human and AI. Nature Medicine. 2019;25:44-56. DOI: 10.1038/s41591-018-0300-7
  12. Kelly CJ, et al. Key challenges for delivering clinical impact with artificial intelligence. BMC Medicine. 2019;17:195. DOI: 10.1186/s12916-019-1426-2
  13. Liu X, et al. Reporting guidelines for clinical trials evaluating artificial intelligence interventions: CONSORT-AI. Nature Medicine. 2020;26:1364-1374. DOI: 10.1038/s41591-020-1034-x
  14. Collins GS, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models. BMJ. 2024;384:e078378. DOI: 10.1136/bmj-2023-078378
  15. Park SH, Han K. Methodologic Guide for Evaluating Clinical Performance and Effect of AI. Radiology. 2018;286:800-809. DOI: 10.1148/radiol.2017171920
  16. Shen J, et al. Artificial Intelligence versus Clinicians in Disease Diagnosis: Systematic Review. JMIR Medical Informatics. 2019;7:e10010. DOI: 10.2196/10010
  17. Vasey B, et al. Reporting guideline for the early-stage clinical evaluation of decision support systems driven by AI: DECIDE-AI. Nature Medicine. 2022;28:924-933. DOI: 10.1038/s41591-022-01772-9
  18. Celi LA, et al. Sources of Bias in AI for Health Care. NEJM AI. 2024;1:AIra2300028. DOI: 10.1056/AIra2300028
  19. Obermeyer Z, et al. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366:447-453. DOI: 10.1126/science.aax2342
  20. Finlayson SG, et al. The Clinician and Dataset Shift in AI. NEJM. 2021;385:283-286. DOI: 10.1056/NEJMc2104626
  21. Esteva A, et al. A guide to deep learning in healthcare. Nature Medicine. 2019;25:24-29. DOI: 10.1038/s41591-018-0316-z
  22. Cabitza F, et al. Unintended Consequences of Machine Learning in Medicine. JAMA. 2017;318:517-518. DOI: 10.1001/jama.2017.7797
  23. Wong TY, Bressler NM. Artificial Intelligence With Deep Learning Technology Looks Into Diabetic Retinopathy Screening. JAMA. 2016;316:2366-2367. DOI: 10.1001/jama.2016.17563
  24. Rajpurkar P, et al. AI in health and medicine. Nature Medicine. 2022;28:31-38. DOI: 10.1038/s41591-021-01614-0
  25. He J, et al. The practical implementation of artificial intelligence technologies in medicine. Nature Medicine. 2019;25:30-36. DOI: 10.1038/s41591-018-0307-0

Article Series: This is Article 35, the final article in the “ML for Medical Diagnosis” research series examining machine learning integration in Ukrainian healthcare.

Keywords: medical AI research agenda, open questions, future research directions, generalization, algorithmic fairness, human-AI collaboration, regulatory science, low-resource healthcare, Ukrainian healthcare, foundation models

Acknowledgments: This research series was conducted as part of PhD studies at Odessa National Polytechnic University, Department of Economic Cybernetics, with support from Stabilarity Hub.

Series Completion: 35 articles published over 12 weeks, covering foundation concepts through practical implementation frameworks for ML-augmented medical diagnosis in Ukrainian healthcare.