Dementia biomarkers, medication safety, frailty, digital health, and clinical AI, with attention to what remains meaningful in real older-adult care.
My work focuses on geriatric medicine, dementia biomarkers, medication safety, frailty, digital health tools, and clinical AI, especially where biological, functional, medication-related, and computational signals meet the practical complexity of older-adult care.
A recurring question is whether a tool, biomarker, study, or intervention remains meaningful when applied to older adults with multimorbidity, polypharmacy, cognitive impairment, functional decline, and caregiver dependence.
In aging medicine, context is not just a footnote or catch phrase. A biomarker, an AI output, a frailty score, a medication list, and a trial endpoint may each describe part of the patient. Taken alone, each can also mistake the part for the whole.
Medicine produces many signals. Blood results, biomarkers, imaging findings, cognitive scores, risk models, wearable data, and now AI-generated outputs all offer a certain kind of precision. The more difficult question is what that precision means in an older patient whose clinical situation does not fit the assumptions behind the measurement.
That is the part of aging medicine I find most interesting. A result may be technically correct and still be hard to interpret. A model may perform well and still answer a question no one at the bedside is asking. A biomarker may separate groups in a research cohort and still become less straightforward in a patient with frailty, vascular disease, inflammation, renal impairment, multiple medications, and a clinical history that is clearer to the caregiver than to the chart.
My work includes interest in cognitive impairment, Alzheimer’s disease, blood-based biomarkers, and circulating miRNA evidence. I am especially interested in how molecular signals should be interpreted in older adults who do not fit neatly into single-disease categories.
Dementia biomarker work is no longer limited to a single marker or single platform. Blood-based Alzheimer’s biomarkers, especially phosphorylated tau assays such as p-tau217, are becoming increasingly relevant to specialist diagnostic pathways. At the same time, miRNA, extracellular vesicles, proteomics, metabolomics, inflammatory markers, neurodegeneration markers such as NfL, and astroglial markers such as GFAP are being studied as parts of a broader biological picture. The geriatric question is not whether these signals are interesting. Many of them are. The harder question is what they mean in a patient with mixed pathology, vascular disease, renal impairment, inflammation, frailty, and medication burden.
Biomarkers can be extremely useful. They can clarify diagnosis, support risk assessment, and help organize research around disease biology. But in older-adult care they rarely arrive alone. They arrive with multimorbidity, frailty, vascular disease, inflammation, medication burden, mixed pathology, and imperfect clinical information. This does not make biomarkers less valuable. It makes interpretation more demanding.
A positive biomarker result in a frail 84-year-old with six concurrent medications carries a different clinical meaning than the same result in a 62-year-old research participant. Translating molecular signals into clinical decisions therefore requires geriatric framing at the interpretation stage, not only at recruitment.
Polypharmacy is one of the central facts of geriatric medicine. It is also one of the easiest to underestimate, partly because a medication list can look quite reasonable when read one drug at a time. Unfortunately, patients do not experience their medication lists one drug at a time.
My work addresses anticholinergic burden, potentially inappropriate prescribing, deprescribing, and drug-related cognitive, functional, and adverse outcomes in older adults. Anticholinergic exposure is a good example of a problem that is both common and often under-recognised. Antihistamines, bladder medications, antidepressants, antipsychotics, sedatives, and other commonly used drugs can accumulate into clinically relevant effects on cognition, falls, delirium risk, and function.
Medication safety is also moving beyond the simple question of whether a drug is “appropriate” or “inappropriate.” Anticholinergic burden scales, CNS-active medication exposure, prescribing cascades, deprescribing strategies, and cumulative medication burden all try to describe a problem patients experience as one body rather than as separate prescriptions. The research question is increasingly whether reducing medication burden changes outcomes that matter: cognition, delirium risk, falls, mobility, hospitalization, independence, and quality of life. Lowering a score is useful. Keeping the patient standing, thinking, and out of hospital is the harder test.
This matters clinically, but it also matters methodologically. In research involving cognition, function, biomarkers, or AI-assisted prediction, medication burden is not just background noise. It can confound the signal, distort the interpretation, and explain part of the outcome that the study may attribute to something else.
Frailty is not the same as age, and it is not simply a longer list of diagnoses. It reflects reduced physiological reserve and increased vulnerability to stressors. In practice, it often explains why two patients with the same diagnosis, the same biomarker, or the same treatment can have very different trajectories.
Function, mobility, falls, neuropsychiatric symptoms, and care dependence are not secondary details in older-adult care. They often determine prognosis, independence, recovery potential, caregiver burden, and the relevance of an intervention more directly than a disease label alone.
The useful question is probably not which frailty measure wins the contest. Fried phenotype, cumulative deficit indices, gait speed, grip strength, sarcopenia measures, mobility, disability-free survival, hospitalization, institutionalization, quality of life, treatment burden, and caregiver-relevant outcomes each capture a different part of the problem. None is magic. Together they remind us that an intervention can improve a disease-specific measure and still fail the older patient if it worsens function, independence, or treatment tolerance.
This is one reason I am interested in how frailty and functional outcomes are measured and used in clinical studies. A study may report a statistically significant result and still leave open the question older patients and families often care about most: did it preserve function, independence, cognition, safety, or quality of life? These are not soft endpoints. They are frequently the endpoints.
Clinical AI brings the same translation problem into sharper view. A model can perform well in a dataset and still be awkward, fragile, or unhelpful in practice. Sometimes the output arrives too late. Sometimes it answers a question no one asked. Sometimes it depends on data that are missing, delayed, unreliable, or entered mainly because someone clicked the least annoying option in the electronic record.
The useful questions in clinical AI are moving beyond “does the model perform well in one dataset?” and toward less glamorous concerns: was there leakage, is the model calibrated, has it been externally and temporally validated, does it drift, does it behave across sites, and does the output change a decision anyone can actually make? For older-adult care, the additional question is whether the validation population includes the people most likely to test the system: frail patients, cognitively impaired patients, multimorbid patients, medication-burdened patients, and patients whose clinical record is more archaeological site than clean spreadsheet.
Some tools are not wrong so much as misplaced. They ask for data that are unavailable, arrive after the decision has already been made, or produce an answer that is accurate but not actionable. In clinical medicine, usefulness is not only a property of the output. It is also a property of timing, workflow, trust, and the person expected to do something with it.
The technical problems are real. Large language models used in documentation, diagnostic support, or evidence synthesis can produce outputs that are fluent, plausible, and wrong. This is a combination medicine has never particularly needed more of. Data leakage can make a model look stronger than it is, especially when information from the future, the outcome label, duplicate patients, or contaminated preprocessing enters the development pipeline. Validation strategy matters because a random split inside one dataset is not the same as performance across time, hospitals, workflows, and patient populations.
Population mismatch is especially important in older-adult care. Many AI systems are developed and tested in populations that are cleaner, younger, less frail, less cognitively impaired, and less medication-burdened than the patients most likely to need support. This does not mean such tools are useless. It means their claims need to be interpreted carefully, and their validation needs to reflect the people in whom they may actually be used.
There are also practical questions that are easy to miss when a project is still living in a slide deck. Can the clinician understand the output? Can it be questioned or overridden? Does it change a decision? Does it create work for someone who already has too much of it? Does it keep working after documentation habits, patient mix, or treatment pathways change? These questions are less glamorous than model architecture, but they often decide whether a tool becomes useful, ignored, or unsafe.
The same concerns apply to clinical trials and aging research more broadly. Older adults are often excluded or simplified through age cut-offs, comorbidity restrictions, polypharmacy limits, cognitive requirements, functional thresholds, and follow-up designs that favor the healthier and more independent. The resulting study population may be internally consistent but poorly representative of the patients who will receive the intervention in practice.
Older-adult study design increasingly has to deal with external validity, pragmatic designs, real-world evidence, patient-centered outcomes, treatment burden, function, quality of life, and endpoints that can handle competing risks rather than politely ignoring them. Hierarchical composite endpoints and functional outcomes are part of this discussion because older patients do not experience survival, toxicity, independence, hospitalization, and cognition as separate academic departments. The difficult question is not only whether an intervention works. It is who it works for, what it costs the patient, and what kind of life is preserved afterward.
For studies involving older adults, endpoint selection, measurement timing, eligibility criteria, frailty assessment, medication burden, cognitive status, and caregiver dependence are not minor technicalities. They shape what the study can honestly claim. Frailty and functional outcomes are often essential complements to biomarker-only, disease-specific, or mortality endpoints. Polypharmacy can affect safety, adherence, cognition, falls, and apparent treatment response.
Small design choices matter. The chosen endpoint, the timing of assessment, the exclusion criteria, the way missing data are handled, the definition of “usual care,” or whether an output is clinically actionable can all change the meaning of a study. These choices rarely announce themselves as major decisions at the time. They prefer to appear later as limitations.
Across these areas, my work returns to a fairly simple question: does the evidence remain meaningful after it leaves the clean setting in which it was generated?
That question connects dementia biomarkers, medication safety, frailty, clinical AI, digital health evaluation, and older-adult trial design. It is the movement from molecular signal to clinical interpretation, from model output to workflow, from trial endpoint to real patient, and from promising idea to evidence that can survive contact with geriatric complexity.