Can AI Predict Disease Before Symptoms Appear?
By Dr. Ande Sriman, M.B.B.S., M.D., Fellowship in Diabetes Mellitus
Consultant Physician and Panel Member of Medicine and Science, 150.fit
One question comes up frequently in clinical practice:
“If I feel completely healthy today, can we know whether my health is beginning to change before I notice symptoms?”
Artificial intelligence is making that question increasingly relevant.
AI models can analyse large amounts of medical information and identify patterns that may be difficult to recognise through individual measurements alone. In research settings, these models have shown the ability to estimate the future risk of certain diseases before a person receives a clinical diagnosis.
But there is an important distinction:
AI can identify patterns associated with future risk. It cannot automatically tell us that someone has, or will definitely develop, a disease.
That distinction between prediction and diagnosis is critical when considering the role of AI in preventive, precision and longevity medicine.
Can AI Really Predict Disease Early?
For some conditions, research suggests that the answer is increasingly yes — within defined clinical and research settings.
Machine-learning models have been studied for predicting conditions including diabetes, cardiovascular disease, neurodegenerative disorders and certain cancers before conventional diagnosis. For example, recent research has evaluated AI models that predicted type 1 diabetes shortly before diagnosis, Alzheimer’s disease several years in advance, and pancreatic cancer using longitudinal medical records before symptoms typically prompted diagnosis. However, predictive performance differs considerably across diseases, populations and models.
So the more useful question is not:
“Can AI predict every disease?”
It is:
“Which health risks can AI estimate reliably enough to improve clinical decision-making?”
That is where the field is moving.
How AI Can Detect Patterns Before Symptoms Appear
For many chronic diseases, biological changes can begin before symptoms become obvious.
Individual laboratory values may still fall within accepted reference ranges while their pattern over time is changing.
AI and machine-learning systems can help analyse these longitudinal patterns across multiple sources of information.
1. Changes in Biomarkers Over Time
A single blood test provides a snapshot.
Repeated measurements can reveal a trajectory.
AI models can analyse changes across metabolic, cardiovascular, inflammatory and other biomarkers to identify combinations or patterns associated with future health outcomes.
This is different from simply flagging whether one result is above or below a reference range.
Recent research in neurodegenerative disease, for example, has found that longitudinal biomarker patterns can sometimes provide more predictive information than baseline measurements alone.
2. Genomic and Multi-Omics Data
An individual’s biology can be studied at several levels.
These may include:
- Genomics
- Epigenomics
- Proteomics
- Metabolomics
- Microbiome data
AI can help analyse relationships across these large datasets that would be difficult to interpret manually.
This does not mean that more testing automatically produces better predictions. The information still needs to be clinically relevant, appropriately validated and interpreted in context.
3. Electronic Health Records
Longitudinal medical records contain information about previous diagnoses, laboratory results, medications, symptoms and healthcare use.
Machine-learning models can analyse these records to identify combinations of factors associated with future disease risk.
Several experimental disease-prediction models have already been developed using large electronic health-record datasets.
4. Wearables and Continuous Health Data
Wearable devices can generate repeated information about variables such as:
- Heart rate
- Heart rate variability
- Sleep
- Physical activity
- Glucose patterns in people using continuous glucose monitors
Instead of evaluating one isolated reading, computational models can analyse how these signals change over time.
The potential value lies in identifying meaningful changes in physiology — not simply generating more health data.
AI Prediction vs Medical Diagnosis
This is one of the most important distinctions to understand. Prediction estimates risk
A predictive model asks:
Based on the available data, how similar is this person's pattern to patterns associated with a future health outcome?
Its output might be a risk estimate, probability or risk category.
Diagnosis determines whether a condition is present
A medical diagnosis requires clinical evaluation using established diagnostic criteria.
Depending on the condition, this might involve:
- Symptoms
- Medical history
- Physical examination
- Blood tests
- Imaging
- Pathology
- Other validated diagnostic tests
An AI-generated risk score therefore does not equal a diagnosis.
A person can have an elevated predicted risk and never develop the disease. Another person can develop disease despite previously being classified as lower risk.
Why Earlier Risk Detection Matters
The potential value of predictive medicine is time.
If clinically validated tools can identify increased risk before disease becomes established, clinicians may have an opportunity to:
- Monitor the person more closely
- Perform appropriate additional testing
- Address modifiable risk factors
- Improve nutrition, physical activity or sleep
- Manage blood pressure, lipids, glucose or other established risks
- Consider medical treatment when clinically indicated
The intervention should be based on validated clinical evidence — not simply on an algorithmic prediction.
That distinction is especially important because early prediction can also create false positives, unnecessary testing, anxiety and overtreatment if the model is not sufficiently accurate.
Where Has AI Disease Prediction Shown Promise?
Research is developing quickly across several areas.
Diabetes and Metabolic Health
AI models can combine information such as medical history, laboratory results, glucose patterns and other metabolic variables to estimate diabetes risk.
Research cited in a recent preventive-AI review included a model using electronic health-record information to predict type 1 diabetes before clinical diagnosis.
For metabolic health more broadly, longitudinal information may eventually help clinicians recognise changes earlier than relying on a single glucose or HbA1c measurement.
Neurodegenerative Disease
Researchers are studying whether combinations of clinical records, blood biomarkers, imaging and other data can identify people at higher risk of conditions such as Alzheimer’s or Parkinson’s disease before obvious neurological symptoms emerge.
These findings are promising, but they should not be interpreted as established population-wide screening tools.
Cardiovascular Disease
Machine-learning models can combine conventional cardiovascular risk factors with imaging, biomarkers, ECG information and longitudinal clinical data.
The aim is to improve risk stratification — identifying which individuals may warrant closer assessment or earlier preventive intervention.
Cancer
AI is being studied extensively in cancer detection and risk prediction.
For example, researchers have developed models using longitudinal health records and blood-test patterns to identify people at increased risk of pancreatic cancer before diagnosis.
However, predicting risk and proving that AI-led screening improves clinical outcomes are two different steps.
What AI Cannot Tell Us Yet
The rapid development of medical AI can make its capabilities sound more certain than they actually are.
A prediction model may perform well in the population on which it was developed but perform differently when used in another population.
Important questions include:
Was the model validated in a separate population?
Does the dataset represent people like the patient being assessed?
How many false positives does the model generate?
Does finding the risk earlier actually improve health outcomes?
Is there an effective intervention available if increased risk is identified?
These questions matter because many disease-prediction models are still at the research or validation stage. Recent analysis has also raised concerns about the quality of datasets used to develop some AI disease-risk models.
AI therefore needs the same scrutiny applied to any other medical tool.
AI and Precision Medicine
This is where AI becomes particularly relevant to precision medicine.Precision medicine considers differences in genetics, environment and lifestyle when making decisions about prevention and treatment. The NIH describes this as moving beyond an exclusively “average patient” approach toward healthcare informed by individual differences. AI can help analyse the large volumes of information that precision medicine may generate.
For example: Genetics + biomarkers + medical history + lifestyle + wearable data
may contain far more information than a clinician could reasonably process simultaneously without computational assistance. AI can help identify relationships within that information.But the algorithm should support clinical judgement, not replace it.
Why Diverse Health Data Matters
One of the biggest challenges in predictive AI is ensuring that models work across different populations.
A model developed largely using data from one ancestry, geography or healthcare system may not perform equally well elsewhere.
This is one reason large research initiatives such as the NIH All of Us Research Program are building diverse datasets containing genomic information, electronic health records, physical measurements, surveys and wearable data.
For precision healthcare to be genuinely personalised, the underlying evidence also needs to represent the people receiving the care.
Looking Ahead
AI has the potential to change how we recognise disease risk.Instead of analysing health only after symptoms appear, increasingly sophisticated computational tools may help clinicians recognise meaningful patterns earlier.But earlier is not automatically better.Prediction becomes valuable only when it is accurate, clinically validated, interpretable and connected to an intervention that can meaningfully improve health. AI should therefore be viewed as an additional clinical instrument — one that can help physicians analyse complex health information and identify patterns that deserve attention.The ultimate goal is not to predict every possible disease.It is to recognise the right risk, early enough to do something useful about it.
References & Context
- NIH Promise of Precision Medicine: National Institutes of Health. The Promise of Precision Medicine. https://www.nih.gov/about-nih/what-we-do/nih-turning-discovery-into-health/promise-precision-medicine
- AI-Driven Biomarker Discovery: AI-driven biomarker discovery: enhancing precision in cancer diagnosis and prognosis. PMC / Nature Publishing Group (2025).
- Machine Learning in Precision Health: AI and Precision Medicine: Paving the Way for Future Treatment. Preprints / Canadian Science Publishing (2024).
- NIH All of Us & Predictive Risk Stratification: Precision Medicine, AI, and the Future of Personalized Health Care. PMC / NIH All of Us Research Program.
