Papers · Paper 2 · preprint, under review

Manifold polarity as the geometric cause of cross patient failure in Riemannian EEG seizure detection

Reframes polarity from binary classification (standard/inverted) to a continuous strength measure. Hardware scaling across CH8, CH12, and CH16 configurations reveals that polarity magnitude depends on circuit depth and coherence budget.

Preprint: Research Square, posted 6 October 2026, CC BY 4.0 · DOI 10.21203/rs.3.rs-10787059/v1

Abstract

Manifold polarity, first identified as a binary property in Paper 1, is more accurately described as a continuous projection of the patient’s covariance shift onto the circuit’s fixed measurement axis. The sign determines polarity direction; the magnitude determines signal strength. Two patient populations emerge: strongly polar (chb03, chb21) whose polarity is stable across configurations, and weakly polar (chb01, chb05, chb07, chb14) whose polarity direction is configuration-dependent.

Hardware validation summary

All results from IBM Torino (Heron r1), 1024 shots, no error mitigation.

ConfigurationChannelsQubitsRaw AUCOracle-Cal AUCN Inverted
CH8 @ 1.95s8170.5240.6373
CH8 @ 20s8170.4040.6056
CH1212250.5220.5654
CH1616330.5120.5120

All AUC values labeled “oracle-calibrated” assume known polarity. CH16 collapses to chance — 246 CZ gates exceed the coherence budget.

Pre-registered prediction outcomes

Honest reporting of pre-registered predictions against actual hardware results:

chb21Confirmed — inverted polarity consistent across CH8@1.95s, CH8@20s, and CH12.
chb14Partial — standard at CH8@1.95s and CH12, but inverted at CH8@20s. Configuration-dependent.
Cross-config consistencyPrediction not met. Only 3/7 subjects show consistent polarity across all configurations. Polarity direction is not purely geometric — it depends on circuit depth and channel selection.

Patient subset rationale

7 patients selected from CHB-MIT: chb01, chb03, chb05, chb07, chb11, chb14, chb21. Selection criteria: sufficient seizure events for LOSO validation, consistent recording montage, and coverage of both polarity populations identified in Paper 1 simulation results.

Status

Posted as a preprint on Research Square on 6 October 2026 (v1, CC BY 4.0), DOI 10.21203/rs.3.rs-10787059/v1. Under review at the journal; first submitted 22 August 2026.

What changed since

The following is the “Sharpened 2026-09” note from the polarity record, verbatim.

Sharpened 2026-09. Within a patient, the seizure shift direction is coherent across seizures (bias-corrected coherence 0.84–1.00 in 22 patients; shape-only median 0.84). Most of the shift — median 83 % — is a uniform power gain that the trace-normalised state cannot hold; the robust, reference-free polarity is the sign of that power change. The inverted set reported in Paper 2 depends on the reference axis, exactly as the fixed-basis mechanism predicts. Details in results/prompt037–039.