← What I found · the finding
Nights that fray before a crash
Partly Worth watching Fragmented sleep as an early sign
What the guide says Awake and restless minutes per night could work as a status readout, a quiet sign the body is struggling. This one comes from the part of her guide she flags as worth building but had not built.
Across the four days before a crash it catches ~7 of 10 crash lead-ups, but fires on about half of ordinary weeks too; when it fires, a crash follows roughly 1 time in 34. Narrowed to the last two nights it fires less often but truer, and clears the chance bar. The one signal here that runs early.
- Catches
- 73%
- of crash lead-ups it fired on (sensitivity)
- Stays quiet
- 49%
- of ordinary days it correctly stayed quiet (specificity)
- When it fires
- 3.0%
- of the days it fires are followed by a crash (precision)
- Lift
- 1.41×
- × more likely than knowing nothing
These describe how the signal behaves in daily life, framed against how rare crashes are, not whether the pattern is real. That verdict is the badge above; a usable-looking number doesn't make a signal that didn't hold real. Tier C (worth watching at best), and with only 29 crashes the margins are wide.
What showed up in the data
I came to this one with no expectation either way, just curious what my body might be doing in the days before a crash. And of everything this watch records, fragmented sleep is the one channel that does something I find genuinely fascinating: it runs ahead of a crash instead of arriving with it. Inside my sleep, the watch counts the minutes I spend awake. On ordinary nights that number is small; in the nights before a crash it runs higher, and the closer the crash gets, the higher it climbs. Across the four days before a crash the signal fires on about 73 percent of crash weeks, against roughly half of ordinary ones. That four-day gap is real in size but just misses the bar that separates a true signal from chance. Narrow the window to the last two nights and it clears the bar cleanly. So the fraying is real, and it is early; it is also modest, a low hit rate on a rare event, the closest thing to a forecast I have rather than a strong one.
How we tested it
It was pre-registered as HA-PG5-precursor before the run: for every crash, was there more within-sleep awake time in the lead-up than on matched ordinary days, judged against a bar fixed in advance (a real-sized gap that also has to clear a permutation null, so a lucky run can't pass for a signal). Pooled across the whole record, all 29 crashes, with a bootstrap confidence interval that is wide at this crash count and reported, not hidden. The four-day window is the primary; the two-night window is the same signal read tighter, not a separate test.
What it means, and what it doesn't
- What it shows
Of everything this watch records, fragmented sleep is the only channel that behaves like a genuine lead-up: it runs above normal in the days before a crash, not just on the crash day. Every other signal I tested reads the crash as it arrives, not before it. That makes this the closest thing to an early sign I have.
- Why it is only 'partly'
The size of the run-up is real, but at the number of crashes I have, the four-day reading falls just short of proving it isn't chance; only the tighter two-night window clears that bar. And even where it clears, the hit rate is low, a weather sign, not an alarm. Honest confidence, not a strong forecast.
- One thread I keep beside the verdict
The fraying is stronger in my earlier years than my recent ones, and those recent years are also my citalopram years. It is tempting to read the medication into that, but with one body I can't separate the drug from everything else that changed at the same time: recovery, steadier pacing, the seasons, or plain chance. So it stays an open question for the driver ledger, not something I can pin on the medication.
Go deeper
The short version is above. These go to the method and the raw research.