Fredrik Strand

Background: The successful design and implementation of artificial intelligence (AI)–driven solutions in health care requires early and continuous multidisciplinary and multiprofessional collaboration. However, diverse disciplinary educational backgrounds, varying languages, and cultural or geographic differences can lead to misunderstandings. To bridge this gap, a structured approach to AI requirements specification can facilitate a shared terminology and a deep mutual understanding among stakeholders, serving both as a guide for technological development and as a means of defining clear pathways for clinical implementation. While technical requirements are well-established in traditional technology development domains, this structured approach remains relatively underused within clinical and social science contexts. Consequently, valuable insights derived from participatory and stakeholder-driven approaches are often overlooked, limiting the relevance and trustworthiness of AI systems in health care settings. Objective: This study presents a methodology for requirements gathering, specification, mapping, and verification, specifically engineered for the complex, multistakeholder environment of clinically applied AI. The methodology was implemented within the specific case of an international multidisciplinary project evaluating an AI-based prediction tool for neoadjuvant chemotherapy treatment response for breast cancer and forming a part of the developed AI validation framework. Methods: The process for AI requirements gathering, specification, and monitoring included 3 iterative rounds of discussion, engaging nearly 150 social, clinical, technical, ethical, and regulatory experts and patients across Europe, South America, North Africa, and Eurasia. It combines established requirements engineering methods (including the MoSCoW [Must have, Should have, Could have, Will not have] framework) with social innovation techniques to ensure inclusivity and contextual relevance. Results: A key finding is the successful development of a structured framework integrating participatory social innovation with formal requirements engineering in an international AI health care setting, through a traceable multisource workflow including clinical, ethical, and regulatory aspects. It is supplemented with an extensive list of 184 actionable consensus-based requirements, categorized by stakeholder group, providing valuable insights for AI researchers in the oncology field with the potential to be transferable to other digital health domains. The requirements align with the fairness, universality, traceability, usability, robustness, and explainability in the AI (FUTURE-AI) framework, ensuring the tool is trustworthy and comprehensive from a multistakeholder perspective, and ensuring comprehensive consideration of all elements of FUTURE-AI. Conclusions: The proposed methodology represents a significant advancement for requirements engineering in digital health by extending traditional technical processes to systematically incorporate nontechnical requirements from diverse global stakeholders. This unified approach is essential for ensuring AI solutions are not only technically robust but also clinically relevant, legally compliant, and socially acceptable.

Three commercial artificial intelligence (AI)–based computer-aided detection systems showed elevated AI scores for screening mammograms up to 10 years before breast cancer diagnosis, demonstrating their potential use in flagging individuals for supplemental screening and earlier detection.

Background and purpose: Low-dose CT screening of high-risk groups has been shown to reduce lung cancer mortality, and the European Council has therefore recommended that member states explore the feasibility and effectiveness of this approach. In this study, we evaluate the implementation of low-dose CT screening for lung cancer in a Swedish female population. Patient/material and methods: Women aged 54–74 years in the Southern Stockholm region were contacted via an electronic questionnaire. Individuals who met the same eligibility criteria as in the NELSON trial (a minimum smoking history of 15 pack-years) were invited. The screening consisted of a baseline low-dose computed tomography (LDCT) scan, with the option for a follow-up scan in cases with intermediate findings. Findings were managed in accordance with modified Fleischner Society guidelines. Results: Between September 2022 and September 2024, 34,580 invitation letters were sent to randomly selected women aged 54–74 years 11,607 individuals (33.4%) completed the questionnaire, whereof 1,106 (10%) met the inclusion criteria. 990 (90%) individuals accepted the invitation and underwent a baseline low-dose CT scan. There were 152 intermediate and 55 positive scans at baseline, and additional eight positive scans at follow-up. Fifteen cases of Lung cancer were found, yielding a positive prediction value (PPV) of 24%. 87% of the lung cancers were in stage IA. Interpretation: Organized lung cancer screening in a Swedish female population proved feasible, demonstrating a good participation rate, and a cancer detection rate consistent with findings from other major screening trials.

BACKGROUND The successful design and implementation of Artificial Intelligence (AI)-driven solutions in healthcare requires early and continuous multidisciplinary and multiprofessional collaboration. However, diverse disciplinary educational backgrounds, varying languages, and cultural or geographic differences can lead to misunderstandings. To bridge this gap, a structured approach to AI requirements specification can facilitate a shared terminology and a deep mutual understanding among stakeholders, serving both as a guide for technological development and as a means of defining clear pathways for clinical implementation. While technical requirements are well-established in traditional technology development domains, this structured approach remains relatively underutilised within clinical and social science contexts. Consequently, valuable insights derived from participatory and stakeholder-driven approaches are often overlooked, limiting the relevance and trustworthiness of AI systems in healthcare settings. OBJECTIVE This study presents a methodology for requirements gathering, specification, mapping, and verification, specifically engineered for the complex, multi-stakeholder environment of clinically-applied AI. The methodology was implemented within the specific case of an international multidisciplinary project evaluating an AI-based prediction tool for neoadjuvant chemotherapy treatment response for breast cancer and forming a part of the developed AI validation framework. METHODS The process for AI requirements gathering, specification, and monitoring included three iterative rounds of discussion, engaging nearly 150 social, clinical, technical, ethical and regulatory experts, and patients, across Europe, South America, North Africa, and Eurasia. It combines established requirements engineering methods (including the MoSCoW framework) with social innovation techniques to ensure inclusivity and contextual relevance. RESULTS A key finding is the successful development of a structured framework that systematically integrates technical feasibility with critical clinical, ethical, and regulatory constraints. It is supplemented with an extensive list of 184 actionable consensus-based requirements, categorised by stakeholder group, providing valuable insights for AI researchers in the oncology field with the potential to be transferable to other digital health domains. The requirements align with the FUTURE-AI framework, ensuring the tool is trustworthy and comprehensive from a multi-stakeholder perspective, ensuring comprehensive consideration of fairness, usability, transparency, universality, robustness, and explainability. CONCLUSIONS The proposed methodology represents a significant advancement for requirements engineering in digital health by extending traditional technical processes to systematically incorporate non-technical requirements from diverse global stakeholders. This unified approach is essential for ensuring AI solutions are not only technically robust but also clinically relevant, legally compliant, and socially acceptable. CLINICALTRIAL Not applicable

Background: The diagnostic process for probable benign breast lesions involves a 1–40% upgrade rate to malignancy when biopsy (cytology and/or histology) is compared with surgery. In a previously conducted clinical randomized trial, we aimed to examine diagnostic discrepancies between prior biopsy results and subsequent vacuum-assisted excision (VAE). Methods: This study is a post hoc analysis of the Swedish VAE randomized trial. Patients were enrolled between November 2019 and August 2022. All patients who underwent a biopsy before VAE were included in this study. Pathology reports from the initial biopsy, VAE, surgical excision, and recurrence were collected. In addition, we conducted clinical follow-up, including imaging, for at least 2 years. Results: The study population included 169 patients with 169 lesions, of whom 71 underwent fine-needle aspiration cytology (FNA), and 126 underwent core-needle biopsy (CNB) before VAE. The diagnostic discrepancy between FNA and VAE was 38% (27/71). The discrepancy between CNB and VAE was 29% (37/126). The upgrade rate to cancer was 7% (5/71) for FNA and 5% (6/126) for CNB. In the CNB group, the highest upgrade rate to cancer occurred in patients with prior atypical ductal hyperplasia (ADH) on CNB (3/12, 25%). Conclusions: The upgrade rate in histopathological diagnosis between prior CNB and VAE was high (15%), and even higher when comparing FNA with VAE (24%). Our findings support avoiding FNA for BI-RADS 3 and 4a lesions and suggest that multi-round VAE may be a safe and effective alternative to surgery for selected cases, particularly those with ADH on CNB.

Background The diagnostic process of probable benign breast lesions involves a 1–40% upgrade rate to malignancy comparing biopsy (cytology and/or histology) to surgery. Within a previous performed clinical randomized trial, we aimed to examine the diagnostic discrepancies between prior biopsy results and subsequent vacuum-assisted excision (VAE). Materials and methods This study is a post hoc analysis of the Swedish VAE randomized trial. Patients were enrolled between November 2019 and August 2022. All patients who underwent a biopsy before VAE were included in this study. Pathology reports from initial biopsy, from VAE, from surgical excision and from recurrence were collected. In addition, we conducted clinical follow-up, including imaging, for at least 2 years. Results The study population included 169 patients with 169 lesions, of which 71 had a fine-needle aspiration cytology (FNA) and 126 had a core-needle biopsy (CNB) before VAE. The diagnostic discrepancy between FNA and VAE was 38% (27/71). The discrepancy between CNB and VAE was 29% (37/126). The upgrade rate to cancer was 7% (5/71) for FNA and 5% (6/126) for CNB. In the CNB group, the highest upgrade rate to cancer occurred in patients with prior atypical ductal hyperplasia (ADH) on CNB (3/12, 25%). Conclusions The upgrade rate in histopathological diagnosis between prior CNB and VAE was high (15%), and even higher when comparing FNA with VAE (24%). Our findings support avoiding FNA for B3 lesions and suggest that multi-round VAE may be a safe and effective alternative to surgery for selected cases, particularly those with ADH on CNB.

Background: Recent prospective studies have shown that AI may be integrated in double-reader settings to increase cancer detection. The ScreenTrustCAD study was conducted at the breast radiology department at the Capio S:t Göran Hospital where AI is now implemented in clinical practice. This study reports on how the hospital prepared by exploring risks from an enterprise risk management perspective, i.e., applying a holistic and proactive perspective, and developed risk mitigation actions. Method: The study was conducted as an integral part of the preparations before implementing AI in a breast imaging department. Collaborative ideation sessions were conducted with personnel at the hospital, either directly or indirectly involved with AI, to identify risks. Two external experts with competencies in cybersecurity, machine learning, and the ethical aspects of AI, were interviewed as a complement. The risks identified were analyzed according to an Enterprise Risk Management framework, adopted for healthcare, that assumes risks to be emerging from eight different domains. Finally, appropriate risk mitigation actions were identified and discussed. Findings: Twenty-three risks were identified covering seven of eight risk domains, in turn generating 51 suggested risk mitigation actions. Not only does the study indicate the emergence of patient safety risks, but it also shows that there are operational, strategic, financial, human capital, legal, and technological risks. The risks with most suggested mitigation actions were ‘Radiographers unable to answer difficult questions from patients’, ‘Increased risk that patient-reported symptoms are missed by the single radiologist’, ‘Increased pressure on the single reader knowing they are the only radiologist to catch a mistake by AI’, and ‘The performance of the AI algorithm might deteriorate’. Conclusion: Before a clinical integration of AI, hospitals should expand, identify, and address risks beyond immediate patient safety by applying comprehensive and proactive risk management.

AI cancer detection models require calibration to attain the desired balance between cancer detection rate (CDR) and false positive rate. In this study, we simulate the impact of six types of mismatches between the calibration population and the clinical target population, by creating purposefully non-representative datasets to calibrate AI for clinical settings. Mismatching the acquisition year between healthy and cancer-diagnosed screening participants led to a distortion in CDR between −3% to +19%. Mismatching age led to a distortion in CDR between −0.2% to +27%. Mismatching breast density distribution led to a distortion in CDR between +1% to 16%. Mismatching mammography vendors lead to a distortion in CDR between −32% to + 33%. Mismatches between calibration population and target clinical population lead to clinically important deviations. It is vital for safe clinical AI integration to ensure that important aspects of the calibration population are representative of the target population.