Personalised Breast Cancer Screening Enters the AI Era

Personalised breast cancer screening is moving from a research ambition towards a practical model of care. In Australia, a newly released review proposes reducing reliance on age alone and tailoring screening to factors such as breast density and family history. That shift could make risk assessment more accurate, but it also creates a demanding FemTech problem: how to combine imaging, clinical history and changing personal data without producing confusion or widening access gaps.

The same question is emerging elsewhere in women’s health. A recent machine-learning study developed diagnostic models for girls with suspected central precocious puberty, a condition in which puberty begins unusually early because the brain’s hormonal signalling system activates ahead of schedule. Together, these developments show where women’s health technology is heading: away from one-size-fits-all thresholds and towards individual probability. The value will depend less on whether an algorithm exists than on whether its inputs, testing and clinical use are trustworthy.

Personalised breast cancer screening needs more than age

Age remains a major breast cancer risk factor and provides a simple basis for population screening. It is not, however, a complete description of risk. Breast density matters because dense tissue is associated with a higher risk of breast cancer and can also make abnormalities harder to see on a mammogram. Family history, inherited genetic variants, previous breast findings and some reproductive or hormonal exposures may also alter risk.

A risk-adapted programme could use these variables to determine when screening starts, how often it occurs and which imaging method is appropriate. A woman at higher estimated risk might be offered earlier or more frequent surveillance, or additional imaging. Someone at lower risk might avoid unnecessary tests and the anxiety, radiation exposure and procedures that can follow false-positive results.

This is not simply a matter of adding breast density to a digital form. Density can be assessed differently across imaging systems and clinicians, while family histories may be incomplete or change as relatives are diagnosed. Risk estimates also need periodic recalculation. The technology must show which inputs drove a recommendation, how certain that recommendation is and what should happen next.

AI for early puberty shows the potential and limits

Central precocious puberty is generally defined by the appearance of secondary sexual characteristics before age eight in girls, although clinical assessment is more nuanced than a single age threshold. Diagnosis can involve growth patterns, pubertal staging, bone age, hormone measurements and, in selected cases, stimulation testing or imaging. Distinguishing true activation of the brain-pituitary-ovarian axis from benign pubertal variants is important because management and follow-up differ.

The machine-learning models reported in Nature address a recognisable clinical challenge: several imperfect signals must be interpreted together. A model may identify combinations that are difficult to capture with a simple rule, potentially supporting referrals and reducing burdensome testing. It may also make specialist expertise more available in settings where paediatric endocrinology services are limited.

Yet development and comparison of models do not establish that they improve care. Performance can fall when an algorithm is used in a different hospital, laboratory system or population. Puberty timing varies with genetics, nutrition, body composition, health conditions and social context. A training dataset that does not represent that variation may produce uneven error rates.

External validation is therefore essential. Useful evidence would include prospective testing, transparent reporting of missed cases and false alarms, calibration across demographic groups, and comparison with clinician-led pathways. The intended role must also be explicit. A decision-support tool that helps a clinician interpret findings carries different risks from an automated system that determines who receives further testing.

Better inputs are the foundation of women’s health AI

Both breast screening and early-puberty assessment expose a central limitation of FemTech: algorithms cannot repair poorly defined or inconsistently collected data. More variables do not automatically produce a better prediction. Measurements must be clinically relevant, standardised and available at the moment a decision is made.

Women’s health data are also unusually time-sensitive. Breast density can change with age, menopause and some hormonal treatments. Family history evolves. Hormone concentrations vary by time of day, pubertal stage, menstrual phase and assay method. A result interpreted without its timing or context may be less informative than it appears.

Longitudinal records could improve this picture by showing direction and pace of change rather than isolated snapshots. For puberty, growth velocity and progression of physical signs may be more informative than one measurement. For breast risk, updated imaging and family history can shift an earlier estimate. Technology is most useful when it preserves this chronology rather than flattening a person into a permanent risk label.

Data governance matters just as much. Information about genetics, reproductive development and breast health is sensitive. Systems need clear consent, limited data access and policies for correction and deletion. They also need a route for human review when the record is wrong or the recommendation does not fit the clinical picture.

Personalised screening can create unequal care

Risk-based care is often presented as inherently more precise, but precision can depend on access to high-quality inputs. Genetic testing, supplementary imaging and specialist assessment are not equally available. Family history may be less informative for people with small families, adoption, estrangement or relatives whose diagnoses were never documented. Rural services may lack the imaging capacity required by a more intensive recommendation.

An algorithm can therefore be accurate in a technical sense while producing an inequitable pathway. If a system identifies higher risk but timely follow-up is unavailable, the digital layer has not solved the practical problem. Implementation studies should measure waiting times, completion of recommended tests, later-stage diagnoses and outcomes across socioeconomic, geographic and ethnic groups, not only predictive accuracy.

Communication is another safety issue. Breast density is a risk factor, not a diagnosis. An elevated model score is not certainty that cancer or progressive puberty is present. Results should be expressed as understandable absolute risk where possible, with a clear explanation of uncertainty and the consequences of each option.

What meaningful personalised care should measure

The next phase of FemTech will not be defined by the number of machine-learning models launched. It will be defined by whether those tools detect important conditions earlier, reduce unnecessary procedures and work fairly outside the institutions where they were built. Clinical oversight remains necessary because models estimate probabilities; they do not account perfectly for preferences, access constraints or every feature of an individual case.

For breast health, records of screening dates, density notifications, family-history changes and hormonal treatment can help preserve context between appointments. For pubertal or hormonal concerns, the timing and progression of symptoms, growth changes and test results are more useful than an isolated entry. These records do not diagnose disease, but they can make clinical conversations more specific.

Changes in breast symptoms, cycle patterns and other hormonal signals only become readable as patterns when they are logged consistently over weeks. This is the kind of pattern Tulsy is built to surface.

Sources: Medical Xpress, Nature.

This content is for informational and educational purposes only. It is not intended to diagnose, treat, cure, or prevent any disease, and should not replace advice from a qualified healthcare professional.