Yes, you have a disease. No, you don’t have a disease.
There is too much complexity to human health for this simplistic, binary method of diagnosis. The phenotypic state of a disease is an ongoing process, while the genotype remains static. Even when it comes to infection, the presence or absence of a pathogen captures little of the entire dynamic biological world inside each of us.
A person’s future development can be predicted with near-certainty by an inherited mutation or chromosomal arrangement, such as cystic fibrosis, Huntington’s disease, and trisomy 21 (Down’s syndrome). Disease development is not an on/off switch. Rather, diseases progress along quantifiable biological spectrums.
The question is, how can that be done in an everyday clinical setting with routine data?
Icahn School of Medicine at Mount Sinai researchers developed a machine learning (ML)-based system that uses routine clinical data to estimate a person’s risk for multiple diseases, potentially revealing hidden illness years before diagnosis.
Published in the Cell Press journal Med, the study demonstrates that SHIMMER stands out due to its use of standard clinical measurements instead of highly specialized data. The practical use of many AI systems in medicine is limited because they require extensive imaging, genomic sequencing, or disease-specific testing. The fact that SHIMMER uses data that is already collected in regular healthcare settings instead could make it easier to implement on a larger scale.
Decoding disease states
Led by Iain S. Forrest, MD, PhD, the research team trained ML models on seven diseases—atrial fibrillation, breast cancer, coronary artery disease (CAD), migraine, rheumatoid arthritis (RA), schizophrenia, and type 2 diabetes (T2D)—using records from the BioMe Biobank in New York and the UK Biobank.
The researchers found that the disease-spectrum scores aligned closely with known risk factors and biological markers. For atrial fibrillation, rising SHIMMER scores tracked with increasing age, obesity, hypertension, and stroke risk. In T2D, the scores rose alongside glucose levels, hemoglobin A1c, triglycerides, and inflammatory markers. CAD scores correlated with smoking, high cholesterol, chronic kidney disease, and established cardiovascular risk calculations.
Importantly, these relationships were not abrupt. Instead, biological changes increased gradually across the spectrum, supporting the idea that disease develops gradually rather than suddenly at diagnosis.
Forrest and colleagues also showed that SHIMMER is capable of identifying disease severity and predicting outcomes. Higher SHIMMER scores were associated with earlier disease onset, more complications, and reduced survival. In CAD, increasing scores tracked with worsening artery blockage, heart failure, arrhythmias, and heart attacks. For RA, higher scores aligned with greater inflammation, worsening anemia, and increased use of immunosuppressive medications.
Some of the strongest findings involved diseases that traditionally lack reliable biomarkers. Schizophrenia, for example, is diagnosed largely through behavioral assessment rather than laboratory testing. However, SHIMMER found obesity, smoking, cannabis use, cardiovascular complications, and poorer survival linked to schizophrenia risk and severity. Spectrum associations were also found in migraine, another condition without biomarkers, including symptom severity and medication use.
Perhaps the most clinically intriguing result was SHIMMER’s ability to flag people who appeared biologically ill despite lacking a formal diagnosis. The researchers identified several people whose routine clinical data strongly resembled known disease profiles despite no diagnosis. One patient with a high T2D SHIMMER score had obesity, hypertension, high glucose, and classic diabetic symptoms but was never diagnosed. Similar patterns appeared for coronary artery disease and rheumatoid arthritis.
In larger undiagnosed populations, elevated SHIMMER scores still predicted abnormal biomarkers and mortality risk. That suggests the system may detect diseases that traditional healthcare pathways miss.
Not a shiny object
The findings challenge the long-standing medical convention of classifying disease in binary terms: either a patient has a condition or does not. The implications could be significant for preventive medicine and population health management. Continuous disease-spectrum scores may help healthcare systems identify high-risk patients earlier and prevent complications.
The researchers envision SHIMMER as an EHR background decision-support tool. Even if they do not meet diagnostic criteria, patients with high scores for diabetes or coronary artery disease may be prioritized for testing, monitoring, or preventive treatment. In that sense, it is complementary to genetic screening.
Still, the work remains a proof of concept rather than a ready-to-deploy clinical tool. The study was retrospective, meaning it analyzed existing records rather than prospectively following patients in real time. The researchers also note concerns about healthcare bias, uneven data quality, and the ethical implications of artificial intelligence in medicine.
Another important limitation is that SHIMMER does not provide definitive diagnoses. Its scores reflect disease burden and biological resemblance rather than calibrated predictions of future illness. Determining how clinicians should act on specific score thresholds will require further study.
The work represents a broader shift in how medicine may conceptualize disease in the AI era. Instead of fixed categories, conditions may increasingly be viewed as dynamic biological trajectories measurable through continuously updated clinical data. If confirmed in future studies, SHIMMER could turn routine medical records into early-warning systems that detect disease before symptoms appear.
The post SHIMMER: Routine Clinical Data Processed Into Scalable Disease State Markers appeared first on Inside Precision Medicine.