Diagnosis is leaving the specialist's office
An untuned general model matched specialist genomic methods, a wearable detected Parkinson's across two continents, and nurses in rural Nepal screened hearts as accurately as cardiologists.
- In a preprint, the general-purpose tabular model TabPFN-3.5 ranked first on 11 of 18 plant traits and had the best median rank against 18 specialised genomic prediction methods, without domain-specific tuning. Self-reported m:…2227
- In a preprint, the wearable-based SAM-PD model detected Parkinson's disease with an AUROC of 0.93 in a US cohort and 0.95 in a European cohort monitored with a different device. Self-reported m:…2306
- In a preprint from rural Nepal, AI-guided handheld echocardiography performed by nurses detected valvular lesions with 95.8% sensitivity and 100% specificity, with near-perfect agreement with cardiologist overread. Self-reported m:…1756
- SAM-PD detected dopaminergic deficit with an AUROC of 0.86, against 0.80 for the standard MDS prodromal Parkinson's risk score. Self-reported m:…2306
A general model beat the specialists
In a preprint posted this week, a general-purpose tabular foundation model called TabPFN-3.5 matched or beat specialised genomic prediction methods without any domain-specific tuning. It ranked first on 11 of 18 plant traits and had the best median rank when compared with 18 specialised methods. In leave-families-out cross-validation it performed on par with GBLUP and RKHS, two of the specialist methods it was compared with. m:…2227
The authors frame the result as evidence for the bitter lesson: general-purpose learners that scale with computation can overtake models built on human knowledge of a domain. The claim is deliberately provocative, and it is worth noting the study's scope: the traits tested were plant traits, and the result is a preprint, not a peer-reviewed finding. But the direction is hard to ignore, because the same pattern is showing up in clinics, not just in genomics. m:…2227
A wearable that spots Parkinson's
A second preprint describes a wearable-based machine learning model, SAM-PD, that detects Parkinson's disease from sleep and activity tracking. It reached an AUROC of 0.93 in a US cohort and 0.95 in an independent European cohort monitored with a different device. It works from sleep and activity metrics: ordinary signals from a device people already wear. m:…2306
The model also detected dopaminergic deficit, a pathological hallmark of the disease, with an AUROC of 0.86, against 0.80 for the standard MDS prodromal Parkinson's risk score, and it matched that score on detecting alpha-synuclein pathology at 0.76. A wearable performing comparably to a specialist risk score is the kind of result that changes who gets screened, because the device is already being worn. m:…2306
A nurse, a handheld scanner and a near-perfect match
A third preprint, a diagnostic accuracy study from rural Nepal, tested AI-guided handheld echocardiography for rheumatic heart disease screening performed by nurse practitioners rather than cardiologists. The nurses' scans detected valvular lesions with 95.8 percent sensitivity and 100 percent specificity, and agreement with cardiologist overread was near-perfect, with a kappa of 0.97. m:…1756
The study is candid about the limits: implementation faced barriers including training requirements, workflow constraints and technical issues. Accuracy in a study is not the same as a screening programme that runs every week. But the result shows the scarce resource, a cardiologist's time, is no longer the bottleneck for this kind of screening. The bottleneck is now organisation, not expertise. m:…1756
What ties them together
The three papers share a shape. In each case a general tool does a specialist's job, and in each case the validation happened outside the lab: a different device in a different continent for the Parkinson's wearable, leave-families-out testing for the genomic model, and nurses rather than cardiologists in rural Nepal. The results are all preprints, so they should be read as signals, not verdicts. But the pattern is consistent enough to matter. m:…2227 m:…2306 m:…1756
The implication for health systems is about access rather than accuracy. Specialist time is the scarcest resource in medicine, and these tools route around it: a model that needs no tuning, a device already on the wrist, a scanner a nurse can carry to a screening camp. The open question in all three cases is deployment, not performance. Accuracy is the easy part. Getting the tool to the people who need it, with the training and workflow to support it, is the work that remains. m:…1756 m:…2306
What is still unresolved
- Whether TabPFN-3.5 generalises beyond plant traits to human genomic prediction.
- Whether SAM-PD changes clinical pathways, or remains a research finding.
- Whether the Nepal screening model scales beyond the study setting given the training and workflow barriers reported.
- All three papers are preprints and have not been peer-reviewed.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.