Your sleeping brain, a veteran's smartwatch, and AI that forgets your face

Your sleep brain waves tonight may predict dementia five years from now

While you were asleep last night, your brain was quietly leaving a mathematical signature that may already distinguish people heading toward dementia from people who are not.

Think of a healthy brain during deep sleep like a well-tuned piano. When you press a key, it produces a rich, complex sound — the fundamental note plus dozens of overtones layering on top of each other. A piano that is going out of tune flattens those overtones. You still hear a note, but it is simpler, more monotone. Researchers at the Study of Osteoporotic Fractures — a large US longitudinal cohort — found something similar in sleep EEG data. They tracked 290 women aged 65 and older, measured the complexity of their brainwave patterns during sleep, and then followed up roughly five years later to see who had developed cognitive impairment. The women whose brains were already trending toward that flatter, simpler signal during deep sleep stages N2 and N3 were significantly more likely to later show signs of dementia. The difference was statistically robust across all four electrode sites tested (p ≤ 0.001). The technical term for what they measured is the Hurst exponent — a number that captures how 'self-similar' a signal is over time, derived using a method called multifractal detrended fluctuation analysis. Simpler brains produce Hurst exponents drifting toward 1.0; healthier ones stay more variable. The catch is real: this study involved only older women, there is no validated clinical test coming from this yet, and the researchers stopped short of reporting how accurately the signal could classify any single individual. What they showed is group-level separation, not a personal prognosis. Think of it as a weather pattern, not a weather forecast for your specific street.

Thirteen veterans on bikes tested whether a smartwatch could help manage PTSD in real time

A cross-country cycling event, thirteen veterans, and a smartwatch that buzzed when their nervous systems were spiking — that is the whole experiment.

Hyperarousal — the state where your nervous system is stuck in high alert, misreading ordinary situations as threats — is one of the hardest parts of PTSD to manage. The trouble is that it often happens before you consciously register it. Researchers in this pilot trial, run during a real-world veterans endurance-cycling event called Project Hero, tried something simple: put a smartwatch on veterans that monitors heart rate and movement continuously, use a machine learning model to detect when arousal is spiking, and then buzz the wearer and ask them to try a brief self-management technique. It is a bit like fitting a boiler with a pressure gauge that tells the homeowner to open a valve before anything overheats — rather than waiting until steam is already escaping. Seven veterans got the smartwatch-plus-coaching app. Three got cycling only. Four stayed home with monitoring but no event. The digital group maintained the anxiety and PTSD symptom improvements that came from the cycling event better after it ended, while the cycling-only group showed a late upswing in hyperarousal scores. Here is the honest catch: thirteen people total, and only seven in the main intervention arm. This is a pilot feasibility study, not a trial you draw policy from. Several participants also reported that when the watch alerted them, they wanted more support than a self-help nudge — the alert created awareness but no clear next step. The signal is genuinely interesting. The sample size means you hold it loosely.

Teaching an AI to screen for depression without learning your gender in the process

An AI trained to detect depression from your voice can also identify your gender with 92% accuracy — a number most people find quietly alarming.

When you speak, your voice carries more than words. It carries pitch patterns, breath timing, tiny hesitations and irregularities that researchers have linked to depression and anxiety. But those same signals also encode your gender, your age, probably your stress level at the time of recording. Useful clinical information and sensitive personal data are baked into the same sound file. A team working on the Androids Corpus — a depression speech dataset — built a system called InfoShield to try to separate them. The analogy is something like trying to extract just the vanilla flavour from a cake batter while leaving everything else — flour, eggs, sugar — intact. It is hard because the ingredients are mixed at a molecular level. InfoShield does this using an approach called information bottleneck: it deliberately compresses the speech signal so it retains what predicts depression but loses what predicts gender and age. The numbers are concrete. Before the protection, an attacker could infer a speaker's gender from the processed signal with 92.6% accuracy. After InfoShield, that dropped to 55.5% — near-chance for a binary category. Age inference fell from 55.7% to 30.3%. Meanwhile, depression classification F1-score actually improved slightly, from 0.723 to 0.784. The catch: this was tested on one dataset only, and no significance testing or confidence intervals were reported. Whether the protection holds across different recording environments, languages, and populations is completely untested. This is a proof of concept that something real is possible — not a deployment-ready product.

Step back and look at what these three stories share. In each one, a passive signal from your body — brainwaves during sleep, a wrist accelerometer during a bike ride, the sound of your voice — is being read to say something about your mental health that you have not explicitly communicated. That is not inherently bad. It is, in fact, kind of remarkable: mental health conditions are notoriously difficult to measure objectively, and these signals offer a route around the subjectivity of self-report. But every step that makes mental health AI more accurate also makes it more revealing. The sleep study shows your brain is legible while you are unconscious. The veterans study shows a machine can spot your stress before you can. The privacy paper exists precisely because the voice study shows an AI for depression detection is also an inadvertent gender classifier. The field is not just getting better at detection — it is accumulating the means to know things about people without asking them. That is worth watching carefully, and not only by researchers.

The veterans cycling study explicitly frames itself as a pilot awaiting a larger powered trial — if the team pursues that, recruitment and design decisions in the next six to twelve months will determine whether this becomes real evidence or stays a promising anecdote. On the sleep-dementia side, the same SOF dataset is being mined by at least two independent groups simultaneously (the deep sleep classification paper also draws on it), which means replication is already underway — just not yet in a new cohort. The open question I would most want answered: do the InfoShield privacy protections survive languages other than English, and does the depression signal itself hold up outside WEIRD research populations?

• Wikipedia: Electroencephalography (EEG) — what it measures and how
• Wikipedia: Post-traumatic stress disorder — symptoms, causes, and treatments
• Wikipedia: Dementia — overview of types including Alzheimer's disease