A new paper in Translational Psychiatry describes a machine-learning approach that flags suicide risk in people with bipolar disorder using a simple blood draw.
That’s not hype. That’s the claim made by the study’s authors themselves.
In the open-access article, Omveer Sharma, Shani Stern, and colleagues explain that they analyzed RNA sequencing from lymphoblastoid cell lines (LCLs) – white blood cells transformed so they can keep dividing.
They looked for differentially expressed genes (DEGs) between high- and low-risk patients and trained models to separate the two groups.
The abstract names several standout genes – LCK, KCNN2, and GRIA1 – and notes pathway enrichment touching primary immunodeficiency, ion channels, and cardiovascular defects.
The team reports “high accuracy” distinguishing low from high suicide risk in bipolar disorder.
A 95% Signal – But What Does It Mean?
Journalist Diana Bletter at The Times of Israel reports the topline number most people will remember: more than 95% accuracy in the model’s predictions. She cites Prof. Shani Stern at the University of Haifa, who led the work with collaborators in Canada and Italy.
According to Bletter’s report, the team used banked samples from six people with bipolar disorder who later died by suicide, alongside 14 additional bipolar patients split between very low risk and higher risk based on history or family history.
They isolated white blood cells, “immortalized” them using Epstein–Barr virus, and performed RNA-seq to read out active gene expression.
Then came the modeling. Bletter writes that the researchers applied repeated cross-validation ~1,700 times, consistently clearing the 95% accuracy bar for identifying who died by suicide.
If that holds in larger, prospective cohorts, it’s hard to overstate the impact.
Bipolar disorder carries a 10–20× higher risk of suicide death, and roughly 40% of patients attempt at some point, Bletter notes, citing Stern. A reliable screening tool would change clinical practice.
Blood Doesn’t Lie? The Surprise in White Cells
One surprise highlighted by Bletter: the signal came from white blood cells, not neurons.
Stern told The Times of Israel that her team expected brain-derived cells to be necessary, yet they still found psychiatric/neuronal signatures in blood.
Stern says the white cells appear to partially mirror what’s happening in the brain, with many neuronal genes expressed in these immune cells.
She adds that her lab is also generating neurons from induced pluripotent stem cells – but stresses that for prediction, the blood test alone may be enough.
The study abstract echoes that theme indirectly through pathway hits and specific ion channel genes such as KCNN2 and GRIA1, which map neatly onto neuronal excitability and synaptic function. That biological plausibility matters; it’s not just a black-box model.
My take: blood as a surrogate is a feature, not a bug. If the biology tracks, clinicians could monitor risk non-invasively over time – exactly the kind of tool psychiatry has lacked.
Promise Meets Peril: The Redacted Concerns
On the YouTube show Redacted, hosts Clayton and Natali Morris walk straight to the uncomfortable questions: privacy, labeling, and downstream misuse.
They worry about a world where a lab value brands someone “high-risk,” and that label leaks to insurers, lenders, employers, or even family court.
They invoke “Gattaca” for a reason. When a predictive score becomes a social sorting tool, rights can erode – especially in mental health, where involuntary holds and forced treatment already exist under certain conditions.
The Morrises also ask a human question doctors will face on day one: what should a person do with a positive result?
It’s not obvious. “Avoid the train tracks” is not a care plan. Real value requires clinical protocols: closer follow-ups, medication adjustments, psychotherapy access, safety planning, and family engagement.
I think Redacted’s skepticism is healthy – not because the science is suspect, but because governance typically lags breakthroughs.
The same test that could save lives could also stigmatize if policies and protections aren’t built in.
How Strong Is the Evidence – Right Now?
The study’s design matters. As described in the abstract and in Bletter’s piece, this is retrospective and involves a small sample with immortalized LCLs, then cross-validation inside that dataset.
That’s a legitimate first step in translational research, but not the finish line. Accuracy estimates – 95% or otherwise – tend to shrink when you move to external validation across new clinics, demographics, and lab pipelines.
To her credit, Bletter quotes Stern emphasizing exactly that: the need for larger, longitudinal studies to confirm reproducibility and stability across populations and real-world settings.
She frames this as a first step toward an integrated risk model blending biomarkers, brain measures, and behavioral data.
Clinically, that’s the right target. No single lab value should decide care; a multimodal model with clear action thresholds and follow-up plans is how you make this safe and useful.
Build Guardrails Before Broad Adoption
First, do no harm needs a modern addendum: do no data harm. If a hospital offers this test, access control, encryption, minimum necessary use, and explicit patient consent must be non-negotiables. Redacted’s concerns about leakage aren’t paranoia; they’re a risk profile.
Second, regulators and payers should move in lockstep with evidence. Tie coverage and use to validated indications (e.g., established bipolar disorder under specialist care), documented benefit (reduced attempts, better engagement), and audit trails for how results guide decisions.
Third, build the care pathway before the test hits scale. A “high-risk” flag should trigger same-week clinician contact, medication review, safety planning, and therapy access – not a purgatory of worry or, worse, punitive responses.
Finally, keep investing in the biology. Stern’s team points to ion channel and immune pathways and specific genes like GRIA1; mechanistic clarity helps design better treatments, not just better predictions.
What Happens Next
If larger, prospective studies reproduce these results, psychiatry gets something it has long needed: a non-invasive, repeatable way to monitor risk in one of the highest-risk populations.
That’s the promise captured both in the study abstract and in Bletter’s reporting.
If governance drifts while the test spreads, the dystopian edge outlined by Clayton and Natali Morris gets sharper. Labels stick. Data leaks. People get boxed in by numbers that were meant to help.
Both futures are possible. The difference will come from what we build around the test – consent, privacy, clinical protocols, and a relentless focus on benefit over harm.
For now, the headline is fair: a blood test may help predict who is most at risk. The fine print is the real work: proving it at scale and protecting the people it is meant to serve.
Gary’s love for adventure and preparedness stems from his background as a former Army medic. Having served in remote locations around the world, he knows the importance of being ready for any situation, whether in the wilderness or urban environments. Gary’s practical medical expertise blends with his passion for outdoor survival, making him an expert in both emergency medical care and rugged, off-the-grid living. He writes to equip readers with the skills needed to stay safe and resilient in any scenario.
