Predictive techniques in medical imaging: opportunities, limitations, and ethical-economic challenges

Abstract

Predictive techniques in medical imaging offer transformative opportunities for early diagnosis and personalized care but raise complex ethical, legal, and economic challenges. This paper explores current advancements, regulatory implications, and risks such as overdiagnosis, alert fatigue, and information overload, emphasizing the urgent need for multidisciplinary frameworks to ensure responsible implementation and sustainable integration into healthcare systems.

The rise of predictive imaging in medicine

In recent years, the medical field has witnessed a significant paradigm shift towards the integration of predictive techniques¹. This evolution began with the advent of genomic profiling, particularly through comprehensive assessments of whole genes, and has progressively expanded into other domains such as omics sciences and advanced imaging-based analyses. These developments not only enhance diagnostic capabilities but also introduce profound implications for medical practice, encompassing ethical, legal, and economic dimensions. Predictive medicine, initially rooted in genetics and genomics, allowed for an unprecedented level of detail in understanding patient predispositions to various diseases. The capability to analyze an entire genome marked the first significant step towards personalized medicine, enabling us to anticipate potential health issues before, if ever, they manifest clinically. This approach has since extended to all other ‘-omics’ sciences, refining disease pattern understanding at a molecular level. Moreover, advancements in radiology, particularly through quantitative and artificial intelligence (AI)-based analysis techniques² together with the adoption of groundbreaking technologies (e.g., Photon Counting CT³)²³ have unlocked even greater potential, offering insights into pathophysiological changes prior to symptom development.

These technologies collectively push the boundaries of how we define and treat patients. The concept of the ‘un-patient’ has emerged⁴, referring to individuals who are not yet symptomatic but may be classified as potential patients based on predictive profiling. This challenges the traditional “reactive nature” of medical practice, and urges a pro-active approach.

Predictive imaging techniques are increasingly integrated into clinical workflows across various modalities. In oncology, radiomics and AI-driven models are used in MRI and CT scans to predict tumor grade, treatment response or risk of recurrence. For example, radiomic signatures extracted from preoperative MRI have shown promise in predicting glioma molecular subtypes, guiding personalized treatment planning. In cardiovascular imaging, coronary CT angiography combined with machine learning can forecast the risk of major adverse cardiac events by analyzing plaque composition, vessel geometry, and perivascular fat attenuation. Similarly, dual-energy CT and spectral photon-counting CT allow quantitative tissue characterization to anticipate outcomes in conditions such as gout or atherosclerosis. In musculoskeletal imaging, predictive models using ultrasound data can identify early degenerative changes in joints, potentially forecasting osteoarthritis progression. These examples demonstrate how imaging is moving beyond static assessment toward dynamic, prognostic roles. By quantifying patterns invisible to the naked eye, predictive imaging is reshaping clinical decision-making, with increasing relevance in triage, treatment selection, and risk stratification.

Ethical and legal challenges

Predicting illness and identifying at-risk individuals introduces multiple challenges. From an ethical perspective, managing predictive information requires careful consideration. Issues such as transparency, explainability, and the appropriate timing and extent of patient disclosure remain under active debate, compounded by the psychological impact on individuals who may never develop the predicted conditions.

In this context, it is worth mentioning that regulatory initiatives such as the European Union’s AI Act, expected to be fully implemented by 2026, are specifically designed to address some of these ethical challenges. The AI Act introduces a risk-based framework that mandates transparency, human oversight, and accountability for high-risk AI systems, which include applications in healthcare. Similarly, in the United States, various regulatory efforts — although more fragmented — aim to reinforce patient rights, algorithmic transparency, and clinical safety through FDA guidance and emerging federal frameworks. These evolving regulations partially respond to concerns regarding liability, data management, and ethical deployment raised in this manuscript. On the other hand, the creation of the European Health Data Space (UE regulation 2025/327) by 2027 and the possibility to use clinical information for secondary use will create more and more opportunities to develop AI integrated data mining from electronic health data, imaging ones included. Public Institution, researchers and even industry will be able to access European health data for research and training purposes. In this view, anonymization – or pseudonymization – together with a ‘general’ consent to use imaging data have been considered a solid wall against unfair use of these data.

Legally, predictive imaging techniques could reshape the landscape of medical liability. Health care providers might face new pressures, as failing to act on predictive insights could lead to legal repercussions. Conversely, the risk of overdiagnosis and overtreatment also looms, potentially leading to unnecessary interventions that may harm patients or cause psychological distress.

Economic sustainability and health system pressure

Economically, the widespread adoption of predictive imaging technologies raises significant concerns regarding the sustainability of health care systems. The costs associated with these advanced diagnostics are not trivial and could lead to a significant increase in healthcare spending if applied broadly: it is imperative to consider the multilayered costs associated with these advanced diagnostic methods. This point must be comprehensively understood and addressed at the regulatory level to determine who is ultimately responsible for bearing these costs—whether it be insurance providers, patients, or government bodies.

Regulatory frameworks must, therefore, ensure economic sustainability and equitable access. It is crucial that these frameworks provide clear guidelines on how costs are allocated, ensuring that they do not disproportionately impact certain groups or lead to financial inequity within the healthcare system. For instance, if insurance providers are mandated to cover predictive diagnostics, conditions must be set to avoid discrimination or excessive premium increases. Alternatively, public health policies may support state subsidies as part of preventive strategies to reduce long-term expenditures. In this scenario, legislative action might be necessary to allocate funds or create incentives for healthcare systems to adopt these technologies.

In the domain of predictive medicine, imaging holds a unique and powerful position. Imaging techniques offer a tangible and visible representation of potential health issues, which can have a profound impact on both patients and physicians. Unlike genetic risks, which may feel abstract and hypothetical, imaging provides concrete evidence of anatomical or pathological changes. This visual component makes the perceived threat more real and immediate, often compelling action. Physicians, too, are influenced by the clarity and immediacy of imaging findings. The visual evidence from imaging studies is difficult to ignore or downplay, making it a critical factor in clinical decision-making. However, there are significant challenges: the increased sensitivity of modern imaging technologies means that incidental findings are more common, leading to potential overdiagnosis and overtreatment. The fear of medico-legal repercussions can drive physicians to pursue aggressive treatments for findings that may never develop into clinically significant issues, that already represents a huge pression to the health system in some countries⁵. Literature suggests that a significant proportion of imaging exams result in incidental findings⁶, raising questions about how to manage these discoveries without causing undue harm or anxiety. Moreover, beyond early disease detection, predictive imaging techniques are increasingly utilized for forecasting disease progression, evaluating therapeutic response, stratifying risk, and guiding individualized treatment planning. These broader applications transform imaging into a dynamic component of patient management rather than a static diagnostic tool.

Another critical issue closely linked to the current context is the overuse of diagnostic imaging, often driven by defensive medicine practices. The unjustified or inappropriate utilization of imaging resources significantly increases the likelihood of incidental findings, many of which may not have clinical relevance. This, in turn, leads to a cascade of further diagnostic procedures, unnecessary interventions, patient anxiety, and increased healthcare costs. Defensive medicine, fueled by fear of litigation and medico-legal consequences, exacerbates this cycle, pushing clinicians towards requesting more tests than clinically necessary. Recognizing and addressing this phenomenon is essential, as only justified and appropriate imaging, aligned with established clinical indications, can mitigate the risks associated with incidentalomas and optimize patient outcomes. Efforts to counteract overutilization should include education campaigns promoting appropriate imaging, and systemic reforms to reduce medico-legal pressures on healthcare professionals.

Technical limitations and AI-driven risks

In addition to defensive practices, the expanding use of AI-driven alert systems in imaging workflows has introduced another layer of complexity. Recent studies have demonstrated that frequent exposure to AI-generated alerts can contribute to “alert fatigue” among radiologists, diminishing their attention to critical notifications and potentially undermining patient safety. For example, a study published in 2024 found that radiologists who reported high-frequency use of AI systems experienced increased levels of burnout and alert desensitization⁷. This phenomenon highlights the need for carefully calibrated AI deployment and alert management strategies to avoid cognitive overload and ensure the clinical utility of predictive technologies.

A particularly complex aspect involves the operational model of advanced AI systems. These algorithms, characterized by multilayered architectures that abstract data beyond human comprehension, introduce significant opacity — a phenomenon often described in the literature as a “black box” problem⁸. This challenges both transparency and explainability in medical decision-making, raising concerns about relying on systems whose internal processes remain obscure. The risk of creating “crystal balls” that produce predictions or truths without the possibility of contradiction poses a problem that must be addressed.

The EU AI Act explicitly addresses the “black box” problem by requiring that high-risk AI systems in healthcare provide meaningful information on their logic and functioning, ensuring a minimum level of explainability for users and authorities. This regulatory approach aims to prevent blind reliance on opaque systems and promotes informed clinical decision-making. In the USA, the FDA’s Good Machine Learning Practice (GMLP) initiative outlines similar principles, although a comprehensive federal law equivalent to the AI Act is still under development.

Additionally, AI models often demonstrate variable performance across different populations. This issue, known as bias or demographic disparity, underscores the necessity for disparity testing and validation across diverse cohorts. Without rigorous disparity testing, predictive models risk perpetuating or amplifying healthcare inequalities, resulting in suboptimal or unsafe outcomes⁹. This discrepancy underscores the need for rigorous validation and continual monitoring of AI systems to ensure their reliability and fairness across diverse patient populations. The opacity and variability of AI outputs necessitate continuous monitoring and recalibration to ensure fairness, reliability, and clinical utility. Ensuring explainability and transparency is critical for maintaining clinician and patient trust in AI-driven recommendations.

A further point is the ability of imaging to make potential diseases visible before they manifest symptoms suggests the need for a careful and considered approach to predictive medicine. The clarity that imaging brings must be balanced with the uncertainty it introduces. Physicians must understand these complexities to ensure that the benefits of predictive imaging are realized without falling into the traps of overdiagnosis and unnecessary interventions¹⁰.

The integration of predictive techniques into medical practice requires a multidisciplinary dialogue involving ethicists, legal experts, economists, and technologists. Developing comprehensive guidelines and frameworks that address the multifaceted challenges of predictive imaging is imperative because the current level of information we can easily extract from a radiological exam is completely different compared to 10 years ago, and with the introduction of the newer AI models this acceleration will grow faster¹¹, and the use of the past models is totally inefficient. These guidelines should not only focus on maximizing the clinical benefits of predictive imaging but also on safeguarding patient rights and ensuring equitable access to these technologies.

The accuracy of predictive models depends heavily on the quality and quantity of data, which must be meticulously managed and analyzed to ensure reliable predictions. This requires a robust technological infrastructure and considerable expertise in data analysis, which may not be uniformly available across all healthcare settings. Ensuring that these resources are accessible and that the workforce is adequately trained is crucial for the successful implementation of predictive imaging techniques. As medicine moves towards a personalized approach¹², the interpretation of imaging data must be tailored to individual patients. The risk associated with anatomical modifications observed in imaging studies should be evaluated based on the specific context of each patient, rather than relying solely on generalized guidelines. This personalized approach can help mitigate the risk of overdiagnosis and ensure that treatments are appropriate and beneficial for the individual patient.

From incidental findings to information saturation

Moreover, this issue integrates with a further and more complex problem related to imaging, which is somewhat peculiar to this discipline: all anatomic structures captured during an exam are subject to scrutiny, even when unrelated to the original clinical query. This concept of untargeted screening was well described by Kwee and colleagues¹³. It indicates that radiological exams comprehensively acquire large volumes and high-density information of the human body, far beyond the requests of a specialist aiming to rule out a suspected clinical pathology. For example, when studying the liver, information about all the abdominal organs, as well as the skeletal, vascular, and muscular components, is acquired. This level of untargeted screening on such a massive scale—considering the billions of diagnostic imaging studies performed worldwide annually—is unprecedented in clinical medicine.

Unlike blood tests, where specific parameters are selectively requested, imaging captures comprehensive, unavoidable anatomical data. This creates a situation where radiologists are ethically and legally obliged to interpret incidental findings, leading to potential overload of healthcare resources.

The extension of this reasoning leads to a question. It is recognized in patient-centred medicine that the secondary prevention of disease—early detection—is beneficial¹⁴. However, it is important to elaborate on this classical medical concept by incorporating some emerging elements. Namely, the multifaceted cost that arises from the increasingly real possibility, with radiological imaging, of detecting dozens, hundreds, or perhaps thousands of anomalies (or variations from expected normality). These anomalies represent a concrete and visible point of alteration, as is evident with imaging, but their definitive impact in terms of disabling biological trajectory is unknown at the detection stage.

The risk is that an enormous “background noise” of data may be generated, potentially obscuring truly significant information (systemic risk of information saturation). The radiologist, and perhaps even the specialist, may be unable to avoid incorporating such visible data due to the ethical and medico-legal risks that may arise. In a situation where increasingly large volumes of anomalies and deviations from the norm are detected (in systems capable of refined profiling and hence personalised rather than categorical medicine), how can this be managed without causing a system crash due to the reporting of everything and the resultant excessive secondary healthcare burden and psychological stress on the patient?

Healthcare sectors, health econometrics, legal medicine, and ethics must find a shared understanding and a shared courage to address this issue. The cost of increasingly personalized healthcare, with rising average life expectancy and the concomitant need for higher levels of healthcare for surviving individuals, represents a fiscal pressure on states that is increasingly complex to sustain¹⁵. Healthcare systems must balance the benefits of early detection against the dangers of information noise and secondary healthcare burdens.

The cost is no longer that of the individual. Translating this to populations, it may become unsustainable. Can the individual specialist bear the burden of this choice? In a 2010 study, published in the New England Journal of Medicine, Epstein and colleagues wrote, “clinicians should withhold information that is likely to overwhelm and distress patients if their having the information would provide no obvious benefit and they don’t ask for it”¹⁶. But is this truly sustainable? How can a radiologist, who must navigate this continuous untargeted screening and encounter numerous findings that “could” be pathological, and in any case “are not normal,” choose not to report this information?

In a study published in the BMJ in 2018, O’Sullivan and colleagues reviewed 20 systematic reviews and provided guidance to help “clinicians and patients weigh up the pros and cons of requesting imaging scans and will help with management decisions after an incidentaloma diagnosis”¹⁷. However, even in this case, what is the cost of choosing not to proceed? It is probably evident that, despite the peculiarities that distinguish each national healthcare system, three key players are missing from this discussion table in addition to the doctor and the patient: the legal expert, those concerned with the cost of the system, and the ethicist.

In light of these reflections, several actionable proposals can be considered:

• For policymakers: Implement regulatory frameworks that mandate transparent reporting standards for AI systems in medical applications, support evidence-based criteria for indications, and promote the creation of multidisciplinary review boards to assess incidental findings management strategies.

• For clinicians: Develop institutional protocols that incorporate decision-support tools for the prioritization and communication of incidental findings, coupled with continuous education programs on the risks of overdiagnosis and alert fatigue. Learning from other medical fields, a shared vocabulary should be proposed. As for genetics (Variant of Uncertain Significance, VUS) or hematology (Monoclonal Gammopathy of Undetermined Significance, MGUS), all the observed variants in imaging procedure should be cumulatively designed as MorphEUS (Morphological Evidence of Undetermined Significance) and clinical advice should be granted, if requested.

• For researchers: Focus on designing AI models that integrate alert prioritization mechanisms, validate performance across diverse populations, and explicitly address explainability and user-centered design to minimize cognitive overload in clinical practice. A repository of MorphEUS should be implemented, so to gain as many other clinical information as possible.

By translating these strategies into practice, it will be possible to better harness the potential of predictive imaging while mitigating associated risks.

In conclusion, while predictive imaging techniques offers transformative opportunities for early detection prognosis and patient management, it demands a critical re-assessment of ethical, legal, and economic paradigms. Through thoughtful regulation and multidisciplinary engagement, we can harness its potential while safeguarding patients and healthcare systems.

Author contributions

L.S. and E.D. conceived the idea for the commentary and wrote the manuscript together. Both authors contributed equally to developing the conceptual framework, critically revised the text, and approved the final version for submission.

Data availability

No datasets were generated or analysed during the current study.

Competing interests

The authors declare no competing interests.