Learnable Phase-Amplitude Coupling Representations for Off- versus On-Levodopa State Decoding in Dyskinetic Hemiparkinsonian Rats
DOI: Pending DOI
Synopsis
Levodopa for Parkinson’s disease can cause dyskinesia and motor fluctuations, motivating interpretable biomarkers for neurostimulation. We developed the Learnable Coupling Network with Temporal attention (LC-Net-T), a phase-amplitude coupling (PAC) decoder for off- versus on-levodopa classification from local field potential (LFP) recordings in hemiparkinsonian rats. It uses fixed phase filters, learnable amplitude filters, PAC maps, and temporal attention. In leave-one-subject-out validation on 10 rats, LC-Net-T achieved the highest Macro-F1 (0.929 ± 0.041 SEM) and ROC-AUC of 0.948 ± 0.030, outperforming power, PAC, and raw-signal baselines. Learned filters and PAC maps revealed corticostriatal coupling patterns, supporting LC-Net-T as a sensing framework for adaptive neuromodulation.
Background
Levodopa remains a core symptomatic therapy for Parkinson’s disease (PD), but chronic treatment can cause dyskinesia and fluctuating motor states (1). Local field potentials (LFPs) access cortico-basal ganglia dynamics and are attractive for adaptive neuromodulation. Beyond band power, phase-amplitude coupling (PAC) captures cross-frequency coordination, in which the phase of a slow rhythm modulates fast activity and may reflect network excitability and synchrony (2). In PD, exaggerated beta-to-gamma or high-frequency PAC in motor cortex and subthalamic recordings relates to motor impairment and is modulated by medication or deep brain stimulation (3,4). In levodopa-induced dyskinesia, abnormal corticostriatal oscillations, ~80-Hz activity, and altered modulation of high-gamma amplitude suggest that PAC may encode medication states (5,6). Thus, PAC is a candidate biomarker for closed-loop neuromodulation. However, conventional PAC pipelines rely on manually selected bands, offline extraction, and fixed statistics, which may miss spectral shifts. Generic electroencephalography (EEG) frameworks transfer poorly to invasive LFPs, where datasets are small, channels sparse, and interpretability critical. To address this gap, we developed the Learnable Coupling Network with Temporal attention (LC-Net-T), asking whether PAC can be represented as a learnable feature space for off- versus on-levodopa decoding in hemiparkinsonian rats. We hypothesized that preserving PAC structure while learning amplitude filters would improve decoding and reveal interpretable corticostriatal patterns.
Methods
We used the public High-resolution Rat Local Field Potential Atlas (LFP-ATLAS; chronic LFPs from healthy and hemiparkinsonian rats) (Fig. A) (7). For baseline-versus-levodopa decoding, region-level LFPs were extracted from caudoputamen (CPu), primary motor area (MOp), and secondary motor area (MOs). Pre-injection recordings (-20 to -10 min) were labeled off-levodopa, and post-injection recordings (45 to 55 min) on-levodopa. The final set included 10 rats and 59 600-s segments, near-balanced by class (30/29). Signals were downsampled to 500 Hz and processed by LC-Net-T (Fig. B). LC-Net-T uses fixed phase filters and learnable amplitude Sinc filters to compute analytic signals and differentiable within-channel mean-vector-length (MVL)-like PAC maps (8) in 20-s windows with 10-s hop. Each map contains real, imaginary, and magnitude components over 9 phase and 4 amplitude bands. A coordinate convolutional neural network (CoordCNN) encoder (9) preserves frequency-bin identity. Window embeddings are integrated by channel attention and bidirectional gated recurrent units (GRUs) with temporal attention (10) to form segment-level logits. Interpretability used learned filters, PAC maps, channel weights, and temporal attention. Leave-one-subject-out (LOSO) validation was compared with power + radial-basis-function support vector machine (RBF-SVM), traditional MVL-PAC + RBF-SVM, raw convolutional neural network (CNN) with temporal-token attention, EEGNet (11)/ShallowConvNet (12), fixed LC-Net-T, and LC-Net-T-Shuffle (amplitude-shuffled control).
Results
Model performance and interpretability analyses are summarized in Fig. C–F. In LOSO validation, LC-Net-T achieved the strongest balanced classification performance for off- versus on-levodopa decoding (Fig. C). Its macro-averaged F1 (Macro-F1; mean ± standard error of the mean, SEM) was 0.929 ± 0.041, higher than LC-Net-T-Fixed (0.905 ± 0.044), traditional MVL-PAC + RBF-SVM (0.843 ± 0.065), LC-Net-T-Shuffle (0.791 ± 0.066), ShallowConvNet (0.701 ± 0.055), Raw CNN with temporal-token attention (0.693 ± 0.068), bandpower + RBF-SVM (0.664 ± 0.069), and EEGNet (0.521 ± 0.068). These results indicate that, by the primary balanced metric, LC-Net-T achieved the most robust fold-wise decoding performance among both PAC-based and raw-signal baselines. The gap from EEG-style baselines suggests that generic raw-signal architectures did not fully exploit the small, sparse, invasive LFP setting.
Area under the receiver operating characteristic curve (ROC-AUC) results provided a complementary view of discriminability. LC-Net-T reached 0.948 ± 0.030, comparable to traditional MVL-PAC + RBF-SVM (0.950 ± 0.034) and LC-Net-T-Fixed (0.946 ± 0.037), and higher than LC-Net-T-Shuffle (0.916 ± 0.044), Raw CNN (0.872 ± 0.066), ShallowConvNet (0.777 ± 0.098), bandpower + RBF-SVM (0.776 ± 0.094), and EEGNet (0.689 ± 0.086). This pattern suggests that PAC-based representations were generally more separable than power-only or raw-signal baselines. However, the advantage of LC-Net-T was most evident in Macro-F1 rather than ROC-AUC, indicating better fold-wise balanced classification after replacing handcrafted classifiers with an end-to-end, attention-based PAC representation.
The ablation results further clarified the contribution of the learnable and coupling-specific components. LC-Net-T-Fixed retained the same PAC-centered architecture but removed high-frequency filter learning; its reduced Macro-F1 indicates that adapting the amplitude bands improved decision-level robustness while preserving the physiological PAC prior. LC-Net-T-Shuffle disrupted amplitude organization before coupling estimation while keeping the overall model family comparable. Its substantial drop in Macro-F1 suggests that LC-Net-T did not rely solely on model capacity, marginal spectral amplitude, or class imbalance, but benefited from structured amplitude-envelope information used to form PAC maps. The remaining above-chance performance of the shuffled control also suggests that medication-state information may exist in non-coupling spectral structure, but coupling-preserving representations contributed to improved balanced decoding.
Learned filter analyses supported this interpretation (Fig. D). The high-frequency filters showed band-dependent center-frequency and bandwidth shifts from the initialized settings, indicating that the model adjusted the PAC amplitude axis to levodopa-state decoding rather than simply reproducing predefined bands. Group-level PAC maps revealed frequency-specific modulation patterns after levodopa administration (Fig. E), with reduced alpha–high-gamma PAC coupling and enhanced theta–high-frequency oscillation (HFO) coupling relative to baseline. These differences were not spatially uniform: region-specific PAC maps localized informative PAC patterns across caudoputamen, primary motor area, and secondary motor area (Fig. F). The PAC maps provide a route for interpreting which region-frequency combinations contributed to the decision rather than treating the decoder as a black-box classifier. Together, the performance, ablation, and visualization results indicate that LC-Net-T combines the discriminative flexibility of learnable neural models with the interpretability of PAC-based electrophysiological biomarkers.
Discussion
The present results suggest that LC-Net-T can serve as a hypothesis-generating decoder for levodopa-related circuit dynamics. The learned high-frequency filters shifted nonuniformly (Fig. D): lower bands moved toward lower centers, higher bands moved upward, and all bands became wider than initialized. This suggests that medication-state information may lie in broad high-frequency envelopes rather than narrow handcrafted gamma/HFO bands. These shifts are consistent with Fig. E, where PAC decreases and increases occupied different amplitude ranges, implying that filter adaptation may track discriminative coupling changes.
The reduced alpha–high-gamma coupling after levodopa is relevant to prior reports in hemiparkinsonian rats linking dyskinesia to ~80-Hz oscillations, corticostriatal synchronization, and reduced low-frequency modulation of ~80-Hz amplitude (5,6). Our finding is compatible with levodopa reorganizing high-gamma activity from phase-locked coupling toward more autonomous oscillations. In Fig. F, MOp appeared to show stronger coupling-related changes than CPu or MOs, suggesting that primary motor cortical LFPs may be sensitive to levodopa-related motor-state transitions. However, this study used 10 rats, region-level LFPs, and two medication windows; thus, the findings should be viewed as off- versus on-levodopa state decoding and biomarker hypotheses, not a validated dyskinesia severity detector. Extending LC-Net-T to cross-region PAC or directed coupling may reveal corticostriatal network mechanisms and strengthen its value for adaptive neurostimulation.
Disclosures
Nothing to disclose by any author(s).
References
1. Turcano P, Mielke MM, Bower JH, Parisi JE, Cutsforth-Gregory JK, Ahlskog JE, Savica R. Levodopa-induced dyskinesia in Parkinson disease: a population-based cohort study. Neurology. 2018;91(24):e2238-e2243.
2. Tort ABL, Komorowski RW, Eichenbaum H, Kopell NJ. Measuring phase-amplitude coupling between neuronal oscillations of different frequencies. J Neurophysiol. 2010;104(2):1195-1210.
3. de Hemptinne C, Ryapolova-Webb ES, Air EL, Garcia PA, Miller KJ, Ojemann JG, Ostrem JL, Galifianakis NB, Starr PA. Exaggerated phase-amplitude coupling in the primary motor cortex in Parkinson disease. Proc Natl Acad Sci U S A. 2013;110(12):4780-4785.
4. van Wijk BCM, Beudel M, Jha A, Oswal A, Foltynie T, Hariz MI, Limousin P, Zrinzo L, Aziz TZ, Green AL, Brown P, Litvak V. Subthalamic nucleus phase-amplitude coupling correlates with motor impairment in Parkinson’s disease. Clin Neurophysiol. 2016;127(4):2010-2019.
5. Halje P, Tamtè M, Richter U, Mohammed M, Cenci MA, Petersson P. Levodopa-induced dyskinesia is strongly associated with resonant cortical oscillations. J Neurosci. 2012;32(47):16541-16551.
6. Belić JJ, Halje P, Richter U, Petersson P, Hellgren Kotaleski J. Untangling cortico-striatal connectivity and cross-frequency coupling in L-DOPA-induced dyskinesia. Front Syst Neurosci. 2016;10:26.
7. Censoni L, Skovgård K, Barrientos SA, Stan TL, Ronaghi A, Cenci MA, et al. High-resolution Rat Local Field Potential Atlas (LFP-ATLAS) (v1). EBRAINS. doi:10.25493/QNVD-9M2.
8. Canolty RT, Edwards E, Dalal SS, Soltani M, Nagarajan SS, Kirsch HE, Berger MS, Barbaro NM, Knight RT. High gamma power is phase-locked to theta oscillations in human neocortex. Science. 2006;313(5793):1626-1628.
9. Liu R, Lehman J, Molino P, Petroski Such F, Frank E, Sergeev A, Yosinski J. An intriguing failing of convolutional neural networks and the CoordConv solution. In: Advances in Neural Information Processing Systems 31. 2018.
10. Cho K, van Merriënboer B, Gulcehre C, Bahdanau D, Bougares F, Schwenk H, Bengio Y. Learning phrase representations using RNN Encoder-Decoder for statistical machine translation. In: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing. 2014:1724-1734.
11. Lawhern VJ, Solon AJ, Waytowich NR, Gordon SM, Hung CP, Lance BJ. EEGNet: a compact convolutional neural network for EEG-based brain-computer interfaces. J Neural Eng. 2018;15(5):056013.
12. Schirrmeister RT, Springenberg JT, Fiederer LDJ, Glasstetter M, Eggensperger K, Tangermann M, Hutter F, Burgard W, Ball T. Deep learning with convolutional neural networks for EEG decoding and visualization. Hum Brain Mapp. 2017;38(11):5391-5420.