← kineticaai.com
Kinetica AI · Cross-Predictor Analysis · Research

Convergence Analysis
Two independent models, one signal.

The ANS Predictor and the Sleep Quality Predictor were trained independently, on different feature candidates, asking different clinical questions. On the 42 days where both have data, they agree 79% of the time — and both independently selected the same feature as their strongest predictor.

42 shared diary days
Sleep vs ANS fatiga models
Agreement: 33/42 days · 78.6%
Prob. correlation: r = 0.66
hrv_rmssd_night_t0 — convergent feature
42Shared days
79%Agreement
r=0.66Prob. correlation
0.78ANS AUC
0.70Sleep AUC
1Shared feature
Why this matters

Independent convergence as methodological validation

In N-of-1 research, replication across independent analyses is the closest substitute for external validation. When two models — trained on different feature sets, optimised for different targets, validated through separate LOO-CV runs — independently converge on the same physiological signal, that convergence is evidence of a real biological relationship rather than overfitting noise.

KEY FINDING

hrv_rmssd_night_t0 — nocturnal RMSSD from the same night — is the feature independently selected by both models as their primary predictor of next-day fatigue. The ANS predictor selected it from a pool of 13 HRV and sleep variables. The Sleep predictor selected it from a different pool of 20 candidates. Both coefficients are negative and of similar magnitude (−0.923 vs −0.823), pointing to the same physiological direction: higher overnight parasympathetic activity reduces fatigue probability the following day.

This is not a trivial finding. The two models use different predictors for the remaining signal — the ANS predictor adds wake minutes and HF power; the Sleep predictor adds lag-1 RMSSD — which explains their different AUC levels. The convergence on the core signal is what matters.

Feature convergence

What each model uses — and where they meet

The highlighted feature (hrv_rmssd_night_t0) appears in both models, independently selected, with nearly identical coefficient sign and magnitude. The remaining features are model-specific and account for the AUC difference.

Technical: Coefficients computed on standardised features (full-dataset fit, not LOO-CV). Both models use L2-regularised logistic regression (C=0.5, class_weight=balanced). Feature selection was greedy forward LOO-CV, max 5 features, ΔAUC ≥ 0.01 stop criterion. Feature pools were different: ANS used 13 HRV+sleep variables; Sleep used 5 nocturnal HRV variables × 4 lags.
Performance comparison

Both models above chance on the same 42 shared days

AUC evaluated on the 42 days where both models have complete feature data (LOO-CV, 1,000× bootstrap CI). The ANS model's wider feature set gives it a 0.08 AUC advantage on this shared subset — but both confidence intervals overlap, and neither model has been prospectively validated.

Technical: Both models re-evaluated via LOO-CV on the shared n=42 days. The ANS predictor's full-dataset AUC (0.83, n=61) is not directly comparable — the shared subset excludes days where sleep_wake_min or hrv_hf_power are missing. Bootstrap CI 1,000 resamples, seed=42.
Day-level agreement

How often do two independent models classify the same day the same way?

Each point is one diary day. The x-axis is the Sleep predictor's estimated fatigue probability; the y-axis is the ANS predictor's. The dashed lines mark each model's Youden threshold (~0.42). Points in the top-right and bottom-left quadrants are days where both models agree. Colour shows actual fatigue level (blue = low-fatigue day, red = high-fatigue day).

Technical: Agreement defined as both models predicting the same class (high/low fatigue) at their respective Youden-optimal thresholds (Sleep thr=0.418, ANS thr=0.418). Pearson correlation of raw predicted probabilities: r=0.659. Agreement does not imply clinical validity — it measures internal consistency between the two approaches.
Interpretation

What the convergence tells us — and what it doesn't

The 79% day-level agreement and r=0.66 probability correlation are higher than chance (expected agreement ≈ 57% given class frequencies), but they are not independent validations — both models use the shared feature hrv_rmssd_night_t0, which drives most of the correlation.

The more meaningful finding is the independent feature selection convergence: two greedy forward searches on different feature pools, both stopping at the same first feature. This is what would survive a strict reproducibility test — not the day-level agreement, which partially reflects the shared predictor.

The 9 disagreement days (21%) are clinically interesting: these are days where autonomic vigilance patterns (ANS model) and nocturnal recovery patterns (Sleep model) send different signals. Analysing these days in detail would be the natural next step for a prospective study.

METHODOLOGICAL IMPLICATION

A combined model using all four features from both predictors (hrv_rmssd_night_t0, hrv_rmssd_night_t1, sleep_wake_min_t2, hrv_hf_power_t0) would be the natural third-layer hypothesis — testing whether the non-overlapping features provide additive predictive value above the shared nocturnal RMSSD signal. This is the planned next analytical step.