BSETS First-Line Philosopher's Stone Report

Cohort And Sleep Data

• Unique scored recordings: 1753
• Unique subjects: 258
• Successfully scored by Philosopher's Stone: 1753 recordings
• ID prefix counts: B=14, F=39, M=199, R=6
• Hypnogram TXT available: 1753 recordings
• Device summary CSV available: 1747 recordings; missing: M001.6, M142.4, M191.1, R002.7, R003.6, R003.7
• Artifact probability file available: 1752 recordings; missing: M136.3.1
• Sleep-period clean fraction available: 1751 recordings
• Sleep-period clean fraction >=20%: 1225 recordings; below threshold or unavailable: 528
• Sleep-period clean fraction >=35%: 822 recordings
• Sleep-period clean fraction >=50%: 538 recordings

Method Notes

• Device summary CSV sidecars contain sleep metrics such as sleep duration, WASO, sleep efficiency, stage percentages, and record_quality_index; this report parses those values but does not compute record_quality_index.
• TXT hypnograms contain 30-second sleep stages. Cohort sleep plots use device-summary metrics when present and hypnogram-derived fallbacks for duration, sleep efficiency, and stage percentages.
• Artifact files contain 2-second channel-level artifact probabilities. Clean fraction is the selected channel's artifact-free fraction within the hypnogram sleep period only, from first sleep epoch through the end of the last sleep epoch.
• 20/75 QC: a recording passes the preprocessing screen when at least 20% of sleep-period artifact segments have artifact probability <=25%. Primary Phi summaries keep all scored sessions; QC-sensitive summaries can use threshold subsets.
• Age-trend sensitivity fits score = intercept + slope x age, with age centered at the cohort median. The intercept is therefore the fitted score at median age, not age zero. Slopes are score units per year.
• Threshold differences compare the fitted score-age line after keeping recordings with clean fraction >=50% versus >=20%; difference uncertainty is estimated by subject-level bootstrap. The QC<20% row is exploratory and shows recordings failing the 20/75 screen.
• Reliability is repeatability from repeated nights: reliability = between-subject variance / (between-subject variance + within-subject variance). No external outcome is used.
• Reliability for averaging k nights is reliability(k) = between-subject variance / (between-subject variance + within-subject variance / k). SEM = sqrt(within-subject variance), and MDC95 = 1.96 x sqrt(2) x SEM.
• Adjacent Bland-Altman plots use consecutive recordings within each subject; delta is later night minus earlier night.
• Bland-Altman delta magnitude is summarized against the cohort score IQR; this helps judge whether adjacent-night deltas are small or large relative to the available score spread.
• Lag-dependent retest stability uses all within-subject pairs grouped by recording-order lag, not only adjacent recordings.
• Subject-centered order drift subtracts each subject's mean score first; the plotted value is therefore the average within-person deviation at each recording order.
• PCA color association uses OLS R2 from colored variable ~ PC1 + PC2, with a permutation p-value from shuffled variable labels.
• The four Phi scores are on different model scales, so their raw means should not be interpreted as directly comparable units.

Subject-level HTML summaries

Data Flow

Counts show recording flow and the selected-channel clean-fraction preprocessing screen.

Cohort Overview

Cohort plots summarize subject-level age/sex/nights and recording-level duration.

Sleep Overview

Sleep metrics use device summary CSV values when present and hypnogram-derived fallbacks where possible.

Sleep Age Sanity

Simple age-sleep sanity checks with one fitted line per metric.

Brain Health Score Distributions

The first column gives quick numeric summaries; IQR is the middle 50% of recordings. The final column is a compact uncolored age trend.

Score Age Scatter Sensitivity

Each row refits the age-score line after applying a different artifact-free EEG threshold. The orange QC<20% row is exploratory. The overlay row shows the >=20% fit in red, >=50% fit in orange, and recordings excluded by the 50% rule as orange x markers.

Score	QC20% slope (95% CI)	QC50% slope (95% CI)	QC20% intercept (95% CI)	QC50% intercept (95% CI)	Slope diff	Slope diff 95% CI	Slope diff p	Intercept diff	Intercept diff 95% CI	Intercept diff p	Interpretation
Brain Health	-0.0077 (-0.0085 to -0.0068)	-0.0084 (-0.0099 to -0.0069)	0.047 (0.036 to 0.059)	0.025 (0.005 to 0.046)	-0.0007	-0.0019 to 0.0005	0.213	-0.022	-0.036 to -0.006	0.007	age slope stable; score level shifts with threshold
Total Cognition	-0.0142 (-0.0148 to -0.0137)	-0.0141 (-0.0151 to -0.0130)	0.562 (0.554 to 0.570)	0.533 (0.519 to 0.547)	0.0002	-0.0007 to 0.0011	0.678	-0.029	-0.041 to -0.016	0.007	age slope stable; score level shifts with threshold
Fluid Cognition	-0.0172 (-0.0178 to -0.0165)	-0.0170 (-0.0182 to -0.0159)	0.606 (0.596 to 0.615)	0.573 (0.557 to 0.589)	0.0001	-0.0008 to 0.0012	0.738	-0.033	-0.045 to -0.019	0.007	age slope stable; score level shifts with threshold
Crystallized Cognition	0.0031 (0.0027 to 0.0035)	0.0030 (0.0024 to 0.0036)	-0.041 (-0.046 to -0.035)	-0.028 (-0.036 to -0.020)	-0.0000	-0.0006 to 0.0005	0.837	0.012	0.006 to 0.021	0.007	age slope stable; score level shifts with threshold

Exploratory QC<20% Fitted Lines

Score	Recordings	Subjects	QC<20% slope (95% CI)	QC<20% intercept (95% CI)	Pearson r
Brain Health	526	167	-0.0061 (-0.0070 to -0.0051)	0.054 (0.039 to 0.068)	-0.49
Total Cognition	526	167	-0.0140 (-0.0147 to -0.0133)	0.609 (0.598 to 0.619)	-0.86
Fluid Cognition	526	167	-0.0168 (-0.0176 to -0.0159)	0.641 (0.629 to 0.654)	-0.87
Crystallized Cognition	526	167	0.0034 (0.0028 to 0.0040)	-0.038 (-0.047 to -0.029)	0.44

The QC20-to-QC50 score-level shift is small and the age slopes are stable; for analyses that require artifact-QC filtering, QC20 remains a reasonable primary filter with QC50 retained as a sensitivity check for absolute score level.

Score Trajectories

Faint gray lines are subjects; the blue line is the median by recording order and the band is the IQR.

Top And Bottom First-Recording Score Trajectories

Green lines are subjects with the top 10 first-recording scores; red lines are the bottom 10, selected separately within each score.

Swimmer Plots

Each panel sorts subjects by that panel's mean score; columns are recording order capped at the first seven recordings. Color limits are the 2.5th and 97.5th percentile within each score.

Lag-Dependent Retest Stability

All within-subject pairs are grouped by recording-order lag. Blue shows median absolute difference; orange shows pairwise retest correlation.

Lag Stability Summary

This compares all within-subject pairs grouped by recording-order lag. A flat median |delta| line suggests that disagreement does not grow strongly as recordings are farther apart.

Score	Lag 1 pairs	Lag 1 median |delta|	Lag 1 r	Largest lag	Largest-lag pairs	Largest-lag median |delta|	Largest-lag r	Median |delta| ratio	Pattern
Brain Health	1495	0.140	0.43	7	25	0.144	0.33	1.03	Similar median |delta| across displayed lags
Total Cognition	1495	0.092	0.78	7	25	0.128	0.65	1.40	Larger median |delta| at longer lag
Fluid Cognition	1495	0.103	0.79	7	25	0.143	0.63	1.39	Larger median |delta| at longer lag
Crystallized Cognition	1495	0.032	0.36	7	25	0.030	0.20	0.96	Similar median |delta| across displayed lags

Subject-Centered Order Drift

Each recording is centered by that subject's mean score. Values far from zero would suggest systematic order effects.

Order Drift Summary

Scores are centered within subject before summarizing recording order. A nonzero mean slope would suggest first-night/order drift rather than stable repeated-night behavior.

Score	Subjects	Mean slope/recording (95% CI)	Median slope/recording	Slope IQR	Mean slope % cohort IQR	t-test p	Sign-test p	Pattern
Brain Health	256	-0.0018 (-0.0060 to 0.0025)	-0.0019	-0.0212 to 0.0179	-0.71	0.411	0.492	CI overlaps zero
Crystallized Cognition	256	0.0039 (0.0009 to 0.0069)	0.0012	-0.0060 to 0.0098	2.21	0.011	0.150	Positive order drift
Fluid Cognition	256	-0.0002 (-0.0038 to 0.0034)	0.0008	-0.0172 to 0.0160	-0.05	0.926	0.662	CI overlaps zero
Total Cognition	256	0.0001 (-0.0029 to 0.0030)	0.0008	-0.0154 to 0.0138	0.03	0.957	0.851	CI overlaps zero

Adjacent Delta Distributions

Adjacent Bland-Altman

Adjacent Bland-Altman uses consecutive within-subject recordings; delta is later minus earlier.

Bland-Altman Scale Reference

Delta size is easiest to judge against the cohort score spread. The ratio below compares the adjacent-night delta IQR with the full cohort score IQR; smaller ratios mean within-subject changes are small relative to between-recording spread.

Score	Adjacent pairs	Median delta	Delta IQR	Cohort score IQR	Delta IQR / cohort IQR	Median |delta|	MDC95
Brain Health	1495	0.004	0.276	0.252	1.10	0.140	0.388
Total Cognition	1495	0.001	0.182	0.297	0.61	0.092	0.268
Fluid Cognition	1495	-0.002	0.204	0.350	0.58	0.103	0.308
Crystallized Cognition	1495	-0.000	0.063	0.176	0.36	0.032	0.204

Reliability

Reliability is estimated from repeated-night variance, not from an external outcome. Reliability = between-subject variance / (between-subject variance + within-subject variance). For the average of k nights, reliability(k) = between / (between + within/k). SEM = sqrt(within-subject variance); MDC95 = 1.96 x sqrt(2) x SEM.

Sleep Association

QC Association

QC correlations are within-subject correlations.

QC Metric Notes

Artifact-Free Fraction is the fraction of sleep-period 2-second segments in the selected F7-O1 channel with artifact probability <=25%; this is the quantity used for the 20/75 screen. Mean, median, and P95 Artifact Probability summarize the model-estimated artifact probability distribution across those segments. F7-O1 Quality Index and Record Quality Index come from device summary sidecars; this report parses them but does not recompute them. Artifact Segments and Record Duration are intentionally omitted from this QC heatmap because they mainly reflect analyzable coverage/duration rather than artifact burden.

Primary Score Correlation

Top Head Correlations

Latent PCA

PCA is an exploratory two-dimensional view of LHL model features, used here to look for obvious quality, score, or cohort structure.

PCA 2D Color Association

This tests whether each colored variable has a linear gradient across the PC1-PC2 map. It is useful for visible structure, but it is not a nonlinear clustering test.

Variable	n	R2 from PC1+PC2	Permutation p	r with PC1	r with PC2
Age	1753	0.620	0.0010	0.35	-0.70
Artifact-Free Fraction	1751	0.136	0.0010	-0.20	-0.31
Brain Health Score	1753	0.552	0.0010	-0.22	0.71
Total Cognition Score	1753	0.870	0.0010	-0.39	0.85
Sex	1753	0.006	0.0060	0.00	0.08
Recording Order	1753	0.004	0.0320	-0.05	-0.04

Inspection Flags

Routine data availability gaps and 20/75 preprocessing attrition are tracked in QC tables, not listed here as red flags.

Output Tables

CSV output tables are omitted from this public report copy; the private analysis workspace remains the source for tabular exports.

Limitations

• F12 age remains unrecoverable. Age `22` is the median-age imputation used for model covariates.
• Six recordings lack the adjacent device summary CSV, so sidecar-only metrics such as WASO and record_quality_index are unavailable for those sessions; hypnogram-derived sleep duration, efficiency, and stage percentages remain available.
• M136.3.1 lacks the 2-second artifact-probability file for the scored segment, so clean-fraction and 20/75 QC status cannot be evaluated from the artifact source for that recording.
• M24.7 has an artifact-probability file, but its TXT hypnogram contains no sleep epochs, so sleep-period clean fraction cannot be computed for that recording.
• The primary BSETS Phi outputs use a wearable frontal-occipital derivation; device/domain differences from the PSG-like training setting remain a validation limitation.