Abstract

Cancer kills roughly 9.7 million people each year and artificial intelligence (AI), genomics and biomarkers are studied and being tested to improve early cancer detection. This review explores their capacities, current limits and the ethical questions these may bring. Genomic testing can improve cancer risk stratification by identifying high-penetrance mutations and individuals in the top 5-10% of polygenic risk scores (PRS), who may have a two to threefold higher risk of developing cancer than the average population. AI algorithms are being developed to suit various types of screening methods and to manage workflows in hospitals. Biomarkers, molecular indicators of a disease, can be used in all stages of preventative care in oncology for risk assessment, early detection and recurrence monitoring. Circulating tumour DNA (ctDNA) is one biomarker capable of detecting cancers and is measurable by liquid biopsy, a less invasive test that analyses bodily fluids.

While these technologies improve early detection, they risk increasing overdiagnosis, entrenching bias and widening inequalities. Giving genetic data to AI is a major privacy concern as patients must know where the data is going, supported by data-protection policies. Furthermore, PRS shows only modest discrimination when added to clinical models (AUC ≈ 0.62), primarily predicts cancer incidence rather than lethality and may perform less accurately in non-European populations. Therefore, this method remains investigational rather than routine. Current cancer screening is only available for certain cancers and can cause false positives, false negatives and overdiagnosis. Its effectiveness is also limited by unequal access to screening, follow-up testing and treatment worldwide. Socioeconomically disadvantaged communities may have less access to these resources, potentially widening existing inequalities. Therefore, equitable access should be prioritised.

Introduction

Cancer remains one of the leading causes of death worldwide, with around 20 million new cases and 9.7 million deaths in 2022 alone (Bray et al., 2024). The foundation for early cancer detection lies in a survival gradient: for breast cancer, five-year survival exceeds 99% when the tumour is localised but falls to roughly 32% once it has metastasised (American Cancer Society, 2026). The potential benefit of earlier detection is therefore considerable.

Yet earlier detection does not necessarily mean better outcomes. Screening can produce false positives and negatives and can identify slow-growing cancers that would never cause harm. Estimates of overdiagnosis in mammography are around 25% (Srivastava et al., 2019), while PSA-detected cancers may be over diagnosed in up to 60% of patients (Ahmed & Dahman, 2026). Additionally, the benefits of screening are not distributed evenly around the globe since access to testing, specialist care and effective treatment varies considerably between populations and healthcare systems.

Recent advances in biomarkers, genomics and artificial intelligence (AI) offer new possibilities for identifying cancer risk and detecting disease earlier. Biomarkers provide molecular information about cancer risk and tumour presence, while liquid biopsies like circulating tumour DNA testing allow genetic material associated with cancer to be detected from the blood. Genomic technologies can identify high-penetrance inherited variants and estimate susceptibility through polygenic risk scores (PRS), while AI analyses medical imaging, genomic information and health records to identify patterns clinicians may overlook. Each of these tools promises more precise, targeted screening, but precision cuts both ways.

These technologies also have significant limitations. Highly sensitive tests may increase overdiagnosis and consequently overtreatment rather than lower mortality. Genomic and AI approaches perform less reliably in populations underrepresented in their training data, and the collection and processing of genomic data raises major concerns about privacy, consent and accountability. Expensive technologies are also more accessible in well-resourced healthcare systems, widening disparities in access to early detection. This paper will explore the extent to which biomarkers, genomics and AI can improve early cancer detection while avoiding overdiagnosis, bias and inequality.

1. Current Limits of Prevention and Screening

1.1 Current screening protocols

Current cancer prevention and screening remain essential for reducing cancer-related mortality, but current methods have important limitations. The latest World Health Organisation data estimates that there were approximately 20.6 million new cases and 9.8 million deaths worldwide in 2024, which is almost one in every six deaths. Lung, breast, colorectal and prostate cancers were among the most diagnosed, yet established population screening programmes are available for only a limited number of cancers. Current protocols generally recommend: mammography for breast cancer; HPV or pap testing for cervical cancer; stool testing or colonoscopy for colorectal cancer; and low-dose CT for certain people at high risk of lung cancer. However, effective routine screening does not yet exist for many cancers, including pancreatic, brain and most ovarian cancers. The effectiveness of available screening also varies according to a person’s age, risk factors and access to follow-up care. This uneven availability means that the benefits of early detection depend heavily on which type of cancer a person develops. 

1.2 Current false positives, negatives and overdiagnosis

Screening can also produce false positive results, in which an abnormality is identified, but cancer is not present. According to a 2024 National Cancer Institute (NCI) report, approximately 10% of mammograms in the United States result in a request for additional testing, but only about 7% of those recalled examinations lead to a cancer diagnosis. More than half of women screened annually for ten years will experience at least one false positive. These results can cause anxiety, financial costs, biopsies and other invasive procedures. False negatives can also create the opposite problem by missing an existing cancer and potentially delaying treatment. Although false positives can affect the patient both physically and psychologically, false negatives can give the patient false reassurance. If the disease is later diagnosed after a false negative, they may experience anxiety, anger, guilt and reduced trust in the healthcare system. Screening can also lead to overdiagnosis by detecting slow-growing cancers that would have never caused symptoms or threatened a person’s life. Therefore, the number of cancers detected is not, by itself, an accurate measure of whether screening improves survival.

1.3 Unequal access

Access is another major limitation. In 2022, cervical cancer caused approximately 350,000 deaths worldwide, and nearly 94% occurred in low- and middle-income countries, even though HPV vaccination and cervical screening can prevent most cases. In 2025, the WHO reported that fewer than 5% of women in many low- and middle-income countries received cervical screening because of cost, weak health systems and logistical barriers. This demonstrates that an effective test cannot improve outcomes if patients cannot access screening, diagnosis and treatment. Therefore, even if biomarkers, genomics and AI improve screening accuracy, they may not reduce cancer-related mortality unless patients can also access diagnosis and treatment.

1.4 Comparing screening methods

The screening protocols described above mainly reflect recommendations used in the United States and other high-income healthcare systems. Screening practices are not the same worldwide. The National Cancer Institute confirms that recommended screening in the United States currently focuses mainly on breast, cervical, colorectal and lung cancers. Even among cancers with established screening, the purpose and limitations of each test are different. Cervical screening methods can help prevent cancer by identifying precancerous cell changes that can then be treated. Mammography and low-dose CT scans mainly detect existing breast or lung cancers at an earlier stage. This demonstrates that screening must be designed for a specific cancer and risk group rather than being applied equally to the entire population.

Screening tests must also balance sensitivity and specificity. Sensitivity describes how successfully a test identifies people who have cancer, while specificity describes how successfully it identifies people who do not. Increasing sensitivity may allow a test to detect more cancers, but it can also produce more false positives if specificity is reduced. This creates a different balance: a test that misses fewer cancers may send more healthy people for unnecessary follow-up procedures. This follow-up burden can include repeated imaging, biopsies, medical costs and months of anxiety. The NCI reports that resolving an abnormal mammogram result can sometimes take one to two years. Prolonged follow-up can delay definitive diagnosis and, where cancer is present, timely treatment. Screening therefore cannot reduce cancer-related mortality unless patients can access timely diagnostic testing and effective treatment after receiving an abnormal result.

2. Genomics and Polygenic Risk

2.1 Genetic architecture of cancer risk

The inherited DNA variation which determines individual cancer risk baseline may have a major influence on screening strategies. Monogenic high-penetrance variants are rare genetic variants in which a single inherited mutation substantially increases lifetime disease risk. These arise from inherited germline mutations in genes that normally suppress tumour growth or repair DNA. Major case studies of monogenic variants include hereditary breast cancer caused by mutations in the BRCA1 and BRCA2 genes, disrupting double-strand DNA repair, resulting in a cumulative lifetime breast cancer risk of 69-72% (Kuchenbaecker et al., 2017).

On the other hand, polygenic cancer risk is the likelihood of developing cancer based on the combined, cumulative effect of thousands of single-nucleotide polymorphisms across a person’s genome. Each common variant has a small, almost negligible effect, but aggregated together can increase the chances of developing cancer. These variants are measured and combined to calculate a polygenic risk score, which shifts the baseline risk up or down across the whole population, stratifying lifetime susceptibility. PRS changes the shape of the risk distribution, which constrains how much extra discrimination biomarkers, AI and screening can achieve.

2.2 How well do polygenic risk scores perform?

Polygenic risk scores serve as risk stratifiers that can estimate a patient’s chance of developing cancer in the future, rather than a diagnostic test that confirms whether a disease is present today. To measure how well a PRS tells future cases apart from healthy individuals, researchers use the Area Under the Curve (AUC) and Harrell’s Concordance Index (C-index). These performance metrics operate on a scale from 0.5 (random guessing) to 1.0 (perfect prediction). Across breast and prostate cancer, adding polygenic risk scores to clinical models typically produces an AUC of approximately 0.62 (Zirpoli et al., 2024; Boyce, Fan, Watson & Murphy, 2013). This indicates modest overall discrimination: when comparing a randomly selected person who develops cancer with one who does not, the model assigns a higher risk score to the person who develops cancer approximately 62% of the time. However, individuals in the top 5-10% of PRS may still have two to three times the average population risk (Jia et al., 2020). These findings measure different aspects of performance: the AUC reflects the model’s overall ability to distinguish between higher- and lower-risk individuals, whereas the two- to threefold figure describes the increased relative risk within a specific high-risk group. PRS shifts risk distributions, predicting who develops cancer more effectively than predicting cancer-specific mortality. Many cases caught through PRS-guided screening may be slow-growing and low-risk tumours that would never have caused any complications if left undetected. Although the metric indicates useful discrimination for identifying who develops cancer, it has a much smaller impact on saving lives and preventing cancer deaths. Because PRS predict cancer incidence more effectively than cancer-specific mortality, many detected cancers may be slow-growing and harmless. Their modest discrimination therefore risks increasing overdiagnosis more than it improves survival, so PRS-guided screening can only deliver benefit when combined with other risk layers and careful screening thresholds.

2.3 Ancestry bias and overdiagnosis

In populations of non-European descent, the model is less effective at separating high- and low-risk individuals from one another, and its calculated risk percentages are less accurate compared to real-world outcomes (Martin et al., 2019). Because foundational genomic databases are built on predominantly European DNA (80-90%) (Fatumo et al., 2022), the statistical models fail to recognise genetic patterns that are not common to that ancestry. As a result, variations in how genes are linked, their frequency and their risk impact across African, Asian or Indigenous genomes make these scores highly inaccurate for non-European populations, with reduced discrimination of their risk of developing cancer and widening existing inequalities (Evans et al., 2022). 

Because PRS predict cancer incidence rather than lethality, using them as a basis for earlier or more intensive screening risks detecting indolent tumours that would never have caused harm. In prostate cancer, for instance, applying PRS to screen and then treat patients who have an increased genetic risk proves to have only modest benefits in terms of lives saved, while increasing false positives and the detection of harmless tumours (Sud, McNeil & Vickers, 2026). Ultimately, by shifting the screening threshold, PRS alter who gets screened and when, reshaping the delicate balance between saving lives and causing harm through overdiagnosis and overtreatment. How this trade-off plays out depends not only on the genetic risk prediction, but also on how these protocols are designed in screening and treatment protocols.

2.4 Clinical readiness and implications

While pathogenic monogenic variants in genes like BRCA1/2 and Lynch syndrome have well-established clinical guidelines to direct patient management, polygenic risk scores are not yet recommended for routine screening, with health systems considering them “to have limited clinical utility” (Trinidad et al., 2025), preferring the use of more familiar techniques. However, major ongoing trials, including WISDOM for breast cancer and BARCODE1 for prostate cancer, are actively evaluating PRS-guided strategies (Fergus et al., 2025; Sud, McNeill & Vickers, 2026). Additionally, EU4Health health assessments support conditional implementation with monitoring for PRS-enhanced breast cancer screening (Pezzullo et al., 2025).

In relation to the limits of current screening, these findings indicate that germline genomics, particularly PRS, only provide a modest ancestry-biased cancer risk probability. Due to this, PRS cannot overcome the essential limits of current screening and prevention strategies, especially due to the frequent overdiagnosis. The calibration problems across different ancestries further highlight that the shortcomings in early detection are not solely ethical or policy failures but rather come from the genetic data and models the system learnt from. Consequently, any integration of PRS into risk-stratified screening, biomarker algorithms or AI-driven pathways will require diverse genomic datasets and carefully monitored pilot implementations.

3. Biomarkers and ctDNA

Biomarkers and circulating tumour DNA can be used for early detection of cancers. Kohaar, Hodges and Srivastava (2024) describe biomarkers as the genomic, epigenomic, proteomic, metabolic and metabolomic molecules that inform risk and diagnosis of a disease. In oncology, these biomarkers serve a variety of purposes such as being indicators of cancer risk, signalling signs for early detection and diagnosis, being used to measure the stage of the cancer, and helping in the preparation of treatment responses. These biomarkers have a role in every stage of preventative medicine, including: germline and other genetic screening methods, allowing for primary preventative care (care done to prevent disease from taking place) by determining at-risk individuals; screening for certain molecules, aiding in early detection of cancer in secondary preventative care (detection of a disease early); and signalling no recurrence of cancer in tertiary care (management of long term effects and prevention of any relapse).

There are a wide variety of biomarkers measured in oncology, with some even sharing similar screening methods. For example, alpha-fetoprotein (AFP) is a biomarker that is used to diagnose liver cancer and is found in the blood, while CD19 and CD22, biomarkers used for the diagnosis of B-cell lymphomas and leukaemia can be found in bone marrow in addition to blood. Furthermore, beyond the biomarkers used for diagnosis, there are others, such as dMMR, that can help identify those at high risk of certain cancer predisposing syndromes. Most notably, screening of genetic biomarkers such as germline testing or checking for ctDNA can help identify individuals at increased risk or with existing disease, as up to 10% of all cancers are caused by inherited genetic changes, which are detectable via screening for genomic biomarkers (as discussed in Section 2). 

Although biomarkers do hold great promise in early diagnosis of cancers, not all are fully understood, leading to cases of overtreatment (see Section 5). For example, PSA is a protein measured to screen for prostate cancer in men over the age of 50 since the 1990s, yet according to Kwon (2019), in many cases the cancers would be benign and the follow-up procedures would only bring risk to the patient. Hence, there is interest in synthetic biomarkers engineered to amplify signs of disease. Examples of these technologies include quantum dots, nanoparticles and genetically encoded synthetic biomarkers. In addition, to prevent invasive screening procedures, new screening methods are being developed, most notably liquid biopsies. According to the US NCI, a liquid biopsy is “a laboratory test done on a sample of blood, urine or other body fluid to look for cancer cells from a tumour or small pieces of DNA, RNA or other molecules released by tumour cells into a person’s body fluid”. It is relatively easy to perform and may reduce some of the sampling limitations associated with taking tissue from a single tumour region. Over the past two decades, this technique has expanded to encompass examinations of blood, urine, cerebrospinal fluid, saliva, amniotic fluid and ascitic fluid, usually measuring CTCs, cfDNA, ctDNA, cfRNA, lncRNAs, miRNA and EVs, among others.

Circulating tumour DNA (ctDNA) refers to the DNA that comes from cancerous cells and tumours composed of usually less than 200 nucleotides in length. These fragments are released into the bloodstream from dead cancerous cells and are not an exact match to an individual’s DNA due to the genetic mutation caused by the cancer. It can be used in detecting and diagnosing a tumour without requiring a sample of tumour tissue which could be in hard-to-access places such as the brain. Beyond early diagnosis, it can also be used in tertiary preventive care by monitoring patients who survived cancer during periods with no symptoms to ensure no recurrence of cancer takes place.

4.  Artificial Intelligence in Oncology

Artificial intelligence is a technology that involves the use of algorithms and mathematical models to analyse and process large and complex datasets. It has the ability to identify patterns and generate predictions. Over the last few years, technologists have adjusted it to interpret and process imaging data, biopsy samples and test results, enabling it to draw conclusions about a diagnosis that can be beneficial to clinicians. This change has the potential for significant impact on early cancer detection as it saves time when looking for a sign of cancerous cells or genetic mutation and reduces the pressure on radiologists (Bulusu et al., 2025). It may also make more diagnostic accuracy as clinicians will have better track of the workflow and less stress to manage (Maleki Varnosfaderani & Forouzanfar, 2024). As stated in Section 1, false negatives and false positives in breast cancer detection are a considerable problem. Pradhan and Sarkhel (2024) studied how a new algorithm developed by Google’s DeepMind, enhanced breast cancer screening. It reduced false positives by 5.7% and false negatives by 9.4%. AI has the potential to be an excellent tool in early cancer detection.

4.1 The algorithm

Alongside detecting early signs of cancer, AI is also able to assist in the organisation of workflows in hospitals, the proposal of specific treatments for patients including the prescription of personalised drugs, and in decisions about which patients should be screened first. Hunter et al. (2022) explained in their paper that the algorithm learns to do so by a method called machine learning (ML). Machine-learning models, including deep-learning (DL) models, learn patterns from training data. Models can be trained using large quantities of information, including survival data, scans, test results and risk factors. Recently, the concept of natural language processing (NLP) has been introduced, and the databases now have the ability to incorporate clinical notes, converting free-text human language into computer-analysable format. This can allow relevant information from patient’s electronic healthcare records (EHRs) to be incorporated into analysis.

In a recent study, Quanyang et al. (2024) examined the role of AI in early lung cancer detection as most cases are detected after the appropriate treatment window. The review states that convolutional neural network (CNN) shows promising performance in oncology applications ranging from tumour identification to prognosis prediction. Furthermore, a new system was introduced in 2022 for cervical cancer in the UK. Data was collated from different hospitals’ databases to make one national database, subsequently providing more information about patterns indicating a risk of cervical cancer (Hunter et al., 2022). It should be noted that the UK is stated as a high-income country; such infrastructure may be less available in lower-resource settings, potentially widening existing inequalities.

4.2 Data and its use

There are a large number of datasets with the potential to aid in detecting cancer, most of which could be analysed by AI. This would allow radiologists, oncologists and GPs to save time and focus on patients at higher risk. The most commonly used methods to locate cancer are imaging modalities such as X-rays, magnetic resonance imaging (MRI), computed tomography (CT), mammography, ultrasounds and positron emission tomography (PET) (Bulusu et al., 2025). Liu et al. (2022) conducted a study to propose an algorithm, called anatomy-aware graph convolutional network (AGN), capable of finding subtle abnormal masses that could indicate a tumour in mammography imaging. After comparing it with traditional benchmarks, they concluded that AGN outperformed the traditional benchmarks. This displays how a well-trained program can implement simple imaging techniques and combine it with other EHRs and patient’s personal data, to determine abnormalities and make a prognosis. This suggests potential for improving early detection; however, greater sensitivity may also contribute to overdiagnosis.

5. Overdiagnosis and Risk Stratified Screening

As cancer detection technology becomes more advanced, it faces a paradox: the better it gets at finding cancer, the more likely it is to find cancer that never needed finding. Overdiagnosis is defined as the correct identification of a cancer or precancerous condition that would never have caused symptoms, harm or death during a person’s lifetime. Unlike a misdiagnosis, which names the wrong disease, overdiagnosis is technically a true finding, except the condition it finds is harmless. However, this finding is rarely treated as “harmless”. The biggest problem is that overdiagnosis triggers overtreatment, and all treatments carry some harm (Welch, 2011). 

Modern screening technologies are now sensitive enough to spot microscopic changes that earlier approaches often missed due to their lower sensitivity.  This leads them to detect slow-growing, or indolent, cellular abnormalities because those are the ones that linger long enough in the body to be detected by screening technologies. However, Cowan et al. (2000) reported that more aggressive cancers tend to appear and progress between screening intervals. Additionally, the most malignant forms of cancer usually produce noticeable symptoms that force a person to seek medical care. Therefore, the evidence suggests that they would be found by diagnostic imaging, biopsies or genetic profiling anyway. Despite this, many tumours have no lethal potential or even risk of progression into invasive cancers.

Real-world examples of overdiagnosis of cancers are surprisingly prevalent in healthcare systems. Srivastava et al. (2019) summarise estimates suggesting that approximately 25% of cancers detected by mammography, 50% of lung cancers detected by chest radiography and/or sputum examination, 13-25% of lung cancers detected by low-dose computed tomography (LDCT), and 50-60% of prostate cancers detected by prostate-specific antigen are over diagnosed. However, it is worth noting that this can vary across different geographical locations, studies and populations, mainly due to differences in screening protocols, follow-up durations, how overdiagnosis itself is measured and other contextual factors (Etzioni et al., 2013). The main issue is that, at the point of diagnosis, it is difficult for doctors to reliably tell which tumours are dangerous and which are indolent. Therefore, detected tumours typically trigger biopsies, surgery or treatment the patient never needed.

Risk-stratified screening is a medical approach that tailors health checks based on an individual’s unique risk level rather than standardising it for the entire population. It involves the use of traditional risk factors (such as family history, lifestyle and age), biomarker levels, genomic risk (e.g., polygenic risk scores) and AI models that combine these into a single risk score. The aim is to estimate the likelihood that an individual will develop a particular illness and adjust the intensity and frequency of the screening accordingly. Therefore, lower risk individuals would undergo screening with less invasive methods, reducing the probability of overdiagnosis, and higher risk individuals would undergo screening more intensively, helping to improve detection amongst them. Some common applications include allocating annual mammograms or MRIs for high-risk women, and less frequent scans for low-risk women. In theory, this is an extremely effective way of lowering over-diagnosis for lower risk groups. However, its real-world effectiveness depends on the accuracy of risk calculation, as a flawed risk model could potentially still over diagnose and worsen the problem that it is designed to mitigate. Additionally, despite increasing scientific interest and policy support, risk-stratified screening has yet to be consistently defined and implemented across cancer sites and countries (Giorgi Rossi et al., 2026). Even so, as genomic, biomarker and AI-based research models continue to advance, risk-stratified screening represents one of the most promising routes towards reducing overdiagnosis without sacrificing early detection.

The cost of overdiagnosis is not one to be taken lightly. The intense psychological toll of a cancer diagnosis can be truly devastating for patients. Approximately 30% of patients may experience cancer-related psychological disorders, including anxiety, depression and post-traumatic stress. Additionally, caregivers of patients with cancer can experience significant emotional, physical and financial stress, which can have a negative impact on their health (Fereidouni et al., 2024). This makes mitigating overdiagnosis increasingly important. Furthermore, overdiagnosis can be a substantial use of healthcare resources, time and effort for both the healthcare system and patients. This burden is worsened by the fact that patients are rarely educated about the harms and benefits of screening procedures, and it may be difficult for clinicians to communicate the concept of overdiagnosis to patients (Srivastava et al., 2019).

As the previous sections show, AI, genomics and biomarkers make early cancer detection increasingly sensitive. However, this sensitivity is what risks intensifying the problem of overdiagnosis. Greater precision in identifying abnormalities does not guarantee better patient outcomes and may instead mean finding more indolent, harmless cases. If overdiagnosis is not addressed alongside technological advances, preventative medicine’s pursuit of earlier detection may come at a cost far bigger than most anticipate.

6. Ethics, Bias, Inequality and Safeguards

Biomarkers, genomics and AI could improve early cancer detection. AI can identify complex and non-intuitive patterns within biomarkers that traditional methods might have missed. This can give doctors information about how patients react to treatment and can help with predicting the progression and response. However, their use also raises ethical and practical challenges, such as algorithmic bias, patient data privacy and equal access. If these issues are properly addressed, they can prove to be extremely useful in the world of healthcare.

Although these technologies are useful, there are concerns that AI might make more mistakes than a doctor when detecting cancer. Evidence shows that AI can perform well on cancer-detection tasks. A study on AI-assisted screening for breast cancer found that AI could successfully identify abnormal scans and reduce workload for medical professionals (Lång et al., 2023). However, the AI’s performance is reliant on the dataset used to train it. AI learns from existing datasets that have been put into its algorithm. Problems can arise if these datasets are not diverse and do not equally represent all populations. If a group is not appropriately represented in the datasets used to train the AI model, that group will be negatively impacted. The model will be less reliable for that group. Research shows that under-representation in genomics could add to already existing issues in healthcare (Corpas et al., 2024). This creates a risk that patients do not receive quality care and that these inequalities are reinforced. A 2023 study done by the NIH showed that biomarkers can vary across different populations (Jiagge et al., 2023). Genetic variation between populations means differences in biomarkers that could potentially affect how accurately they can predict disease. Verifying that AI is trained using diverse datasets and checked across different populations helps ensure that inequalities will not be perpetuated (Alum, 2025).

There is also concern surrounding patient privacy in relation to the use of genomics and AI because these systems need medical and genomic data. Genetic data is extremely sensitive because it is unique to the patient and can reveal lots of information about them and potentially their biological relatives. This is why the patient must know where their data is going and how it is being used, along with data-protection policies (Kohan, 2026).

Another concern is equal access to these resources. Advanced technologies like AI and genetic testing can become expensive quickly as they often require specialists and facilities. This means that patients in well-resourced healthcare settings are more likely to have access to these technologies, while underprivileged communities have less access. This creates a gap where people with less healthcare access are less likely to benefit from these technologies. We must keep this in mind when distributing these new technologies and prioritise equal access (Dusic et al., 2022).

These risks can be managed and reduced through safeguards. AI systems need to be tested using diverse datasets and monitored for errors when working with scans from diverse groups of people. Medical professionals must oversee and review AI-generated outputs before making clinical decisions, ensuring that patients are receiving quality care without computer error. Patients must be clearly informed about how their data is used, as well as provided with data-protection policies. Finally, healthcare systems must always prioritise access and equality across all populations for these new technologies.

Overall, biomarkers, genomics and AI are useful tools. They have the potential to improve early cancer detection and help patients globally. However, it depends heavily on how they are managed. We must have proper protocols and procedures in place so that we can ensure they benefit everyone.

Conclusion

In conclusion, biomarkers, genomics and artificial intelligence can improve early cancer detection, but their benefits must be balanced against the risks of overdiagnosis, bias and inequality. Together they can be used in all stages of preventative care in oncology, including the use of biomarkers in early cancer detection, genomics in identifying high-risk individuals, and AI in improving screening methods and sorting through information quicker.

However, they introduce concerns regarding overdiagnosis, bias and inequality: biomarkers such as PSA risk overdiagnosis and may have invasive measuring procedures; genomic screening primarily predicts incidence rather than lethality and may perform less accurately in non-European populations; and AI cannot be implemented into every healthcare system as it is not yet completely accessible, possibly causing widening inequality. In addition, the use of genomics and AI can raise concerns regarding privacy and patient confidentiality with the sensitive information in the databases. Therefore, developments in these fields should focus on minimising the flaws prevalent in these technologies. For example, focusing on non-invasive procedures for measuring biomarkers, such as using liquid biopsies and measuring ctDNA, developing synthetic biomarkers to avoid overdiagnosis, accounting for different population groups for PRS, or addressing accessibility issues for AI to prevent inequality.

Overall, these technologies can significantly improve cancer detection, but only when improved detection comes with safeguards that ensure the cancers being detected are the ones most likely to affect a patient’s health.

Ahmed, A.E. and Dahman, B. (2026) Beyond baseline: Serial PSA measurements and risk stratification for prostate cancer, The Prostate, 86(4), pp. 505–507.

Alum, E.U. (2025). AI-driven biomarker discovery: enhancing precision in cancer diagnosis and prognosis, Discover Oncology, 16(1).

American Cancer Society (2026) Survival rates for breast cancer, ACS [online]. <https://www.cancer.org/cancer/types/breast-cancer/understanding-a-breast-cancer-diagnosis/breast-cancer-survival-rates.html>

Boyce, S., Fan, Y., Watson, R.W. and Murphy, T.B. (2013) Evaluation of prediction models for the staging of prostate cancer, BMC Medical Informatics and Decision Making, 13, 126.

Bray, F., Laversanne, M., Sung, H., Ferlay, J., Siegel, R.L., Soerjomataram, I. and Jemal, A. (2024) Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries, CA: A Cancer Journal for Clinicians, 74(3), pp. 229–263.

Bulusu, G., Vidyasagar, K.E.C., Mudigonda, M. et al. (2025) Cancer detection using artificial intelligence: a paradigm in early diagnosis, Archives of Computational Methods in Engineering, 32, pp. 2365–2403.

Corpas, M., Pius, M., Poburennaya, M., Guio, H., Dwek, M., Nagaraj, S., Lopez-Correa, C., Popejoy, A. and Fatumo, S. (2024) Bridging genomics’ greatest challenge: the diversity gap, Cell Genomics, 5(1), 100724.

Cowan, W.K., Angus, B., Gray, J.C., Lunt, L.G. and al-Tamimi, S.R. (2000) A study of interval breast cancer within the NHS breast screening programme, Journal of Clinical Pathology, 53(2), pp. 140–146.

Dusic, E., Bowen, D.J., Bennett, R., Cain, K.C., Theoryn, T., Velasquez, M., Swisher, E., Brant, J.M., Shirts, B. and Wang, C. (2022) Socioeconomic status and interest in genetic testing in a US-based sample, Healthcare, 10(5), 880.

Etzioni, R., Gulati, R., Mallinger, L. and Mandelblatt, J. (2013) Influence of study features and methods on overdiagnosis estimates in breast and prostate cancer screening, Annals of Internal Medicine, 158(11), pp. 831–838.

Evans, D.G., van Veen, E.M., Byers, H., Roberts, E., Howell, A., Howell, S.J., Harkness, E.F., Brentnall, A., Cuzick, J. and Newman, W.G. (2022) The importance of ethnicity: Are breast cancer polygenic risk scores ready for women who are not of White European origin?, International Journal of Cancer, 150(1), pp. 73–79.

Fatumo, S., Chikowore, T., Choudhury, A., Ayub, M., Martin, A.R. and Kuchenbaecker, K. (2022) A roadmap to increase diversity in genomic studies, Nature Medicine, 28(2), pp. 243–250.

Fereidouni, Z., Dehghan Abnavi, S., Ghanbari, Z., Gashmard, R., Zarepour, F., Khalili Samani, N. and Ghasemi, A. (2024) The impact of cancer on mental health and the importance of supportive services: mental health and supportive services for patients with cancer, Galen Medical Journal, 13, e3327.

Fergus, K.B., Heise, R.S., Madlensky, L. et al. (2025) Integrating breast cancer polygenic risk scores at scale in the WISDOM Study: a national randomized screening trial, Genome Medicine, 17(1), 97.

Giorgi Rossi, P., Senore, C., Giordano, L. et al. (2026) Risk-stratified cancer screening: a consensus-based conceptual framework from EUCanScreen, European Journal of Cancer, 240, 116746.

Hunter, B., Hindocha, S. and Lee, R.W. (2022) The role of artificial intelligence in early cancer diagnosis, Cancers, 14(6), 1524.

Jia, G., Lu, Y., Wen, W., Long, J., Liu, Y., Tao, R., Li, B., Denny, J.C., Shu, X.O. and Zheng, W. (2020) Evaluating the utility of polygenic risk scores in identifying high-risk individuals for eight common cancers, JNCI Cancer Spectrum, 4(3),pkaa021.

Jiagge, E., Jin, D.X., Newberg, J.Y. et al. (2023) Tumor sequencing of African ancestry reveals differences in clinically relevant alterations across common cancers, Cancer Cell, 41(11), pp. 1963–1971.

Jooma, S., Hahn, L.A. and Bonham, V.L. (2019) Defining and achieving health equity in genomic medicine, Ethnicity & Disease, 29(Suppl 1), pp. 173–178.

Kohan, L. (2026) Artificial intelligence and genomic data privacy: balancing innovation with security, Molecular Biology Research Communications, 15(1), 1.

Kohaar, I., Hodges, N.A. and Srivastava, S. (2024) Biomarkers in cancer screening: promises and challenges in cancer early detection, Hematology/Oncology Clinics of North America, 38(4), pp. 869–888.

Kuchenbaecker, K.B., Hopper, J.L., Barnes, D.R. et al. (2017) Risks of breast, ovarian, and contralateral breast cancer for BRCA1 and BRCA2 mutation carriers, JAMA, 317(23), pp. 2402–2416.

Kwon, E.J. (2019) ‘Synthetic biomarkers for cancer detection and diagnosis’ in National Academy of Engineering, Frontiers of Engineering: Reports on Leading-Edge Engineering from the 2018 Symposium (Washington, DC: National Academies Press).

Lång, K., Josefsson, V., Larsson, A.-M. et al. (2023) Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI), Lancet Oncology, 24(8), pp. 936–944.

Liu, Y., Zhang, F., Chen, C., Wang, S., Wang, Y. and Yu, Y. (2022) Act like a radiologist: towards reliable multi-view correspondence reasoning for mammogram mass detection, IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10), pp. 5947–5961.

Maleki Varnosfaderani, S. and Forouzanfar, M. (2024) The role of AI in hospitals and clinics: transforming healthcare in the 21st century, Bioengineering, 11(4), p. 337.

Martin, A.R., Kanai, M., Kamatani, Y., Okada, Y., Neale, B.M. and Daly, M.J. (2019) Clinical use of current polygenic risk scores may exacerbate health disparities, Nature Genetics, 51(4), pp. 584–591.

National Cancer Institute (n.d.) Liquid biopsy, NCI Dictionary of Cancer Terms [online]. <https://www.cancer.gov/publications/dictionaries/cancer-terms/def/liquid-biopsy>

National Library of Medicine (2021) What is circulating tumor DNA and how is it used to diagnose and manage cancer?, MedlinePlus Genetics [online]. <https://medlineplus.gov/genetics/understanding/testing/circulatingtumordna/>

Pezzullo, A.M., Taha, A., Pascucci, E., Pastorino, R. and Boccia, S. (2025) Polygenic risk scores in breast cancer screening: recommendations from the Can.Heal project, European Journal of Public Health, 35(Suppl 4), ckaf161.1169.

Pradhan, U. and Sarkhel, D. (2024) Future perspective on AI in breast cancer detection: a mini review, Journal of Artificial Intelligence and Robotics, 1(3), pp. 1–5.

Quanyang, W., Yao, H., Sicong, W. et al. (2024) Artificial intelligence in lung cancer screening: detection, classification, prediction, and prognosis, Cancer Medicine, 13(7).

ScienceInsights (2026) How many people die of cancer every year, worldwide?, ScienceInsights [online]. <https://scienceinsights.org/how-many-people-die-of-cancer-every-year-worldwide/>

Srivastava, S., Koay, E.J., Borowsky, A.D., De Marzo, A.M., Ghosh, S., Wagner, P.D. and Kramer, B.S. (2019) Cancer overdiagnosis: a biological challenge and clinical dilemma, Nature Reviews Cancer, 19(6), pp. 349–358.

Sud, A., McNeill, A. and Vickers, A.J. (2026) Comparison of results from the BARCODE1 study and contemporary prostate cancer screening trials, European Urology Oncology, 9(3), pp. 524–528.

Trinidad, S.B., Fullerton, S.M., Cohn, B., Crosslin, D.R. and Jarvik, G.P. (2025) Polygenic risk scores in the clinic: health-system leaders and primary care providers weigh in, Genetics in Medicine, 27(11), 101575.

Welch, H.G. (2011) Overdiagnosed: making people sick in the pursuit of health, Journal of Clinical Investigation, 121(8), 2954.

Zirpoli, G.R., Pfeiffer, R.M., Bertrand, K.A., Huo, D., Lunetta, K.L. and Palmer, J.R. (2024) Addition of polygenic risk score to a risk calculator for prediction of breast cancer in US Black women, Breast Cancer Research, 26(1), 2.