FRAME-C: A knowledge-augmented deep learning pipeline for classifying multi-electrode array electrophysiological signals

Nisal Ranasinghe; Dzung Do-Ha; Chathura J. Gamage; Simon Maksour; Tamasha Malepathirana; Sachith Seneviratne; Lezanne Ooi; Saman Halgamuge

eess.SPMay 18, 2025

FRAME-C: A knowledge-augmented deep learning pipeline for classifying multi-electrode array electrophysiological signals

Authors: Nisal Ranasinghe, Dzung Do-Ha, Chathura J. Gamage, Simon Maksour, Tamasha Malepathirana, Sachith Seneviratne, Lezanne Ooi, Saman Halgamuge

Organizations: AI, Optimization and Pattern Recognition Research Group, Dept. of Mechanical Eng., University of Melbourne, Australia · Illawarra Health and Medical Research Institute, School of Biological Sciences, Faculty of Science, Medicine and Health, University of Wollongong, Australia

Abstract

Amyotrophic lateral sclerosis (ALS) is a fatal neurodegenerative disorder characterized by motor neuron degeneration, with changes to neuronal excitability being an early phenotype for ALS identified in patients and animal models. Recent advancements in induced pluripotent stem cell (iPSC) technology have enabled the generation of human iPSC-derived neuronal cultures, which, when combined with multi-electrode array (MEA) electrophysiology, provide rich spatial and temporal electrophysiological data. Traditionally, MEA data is analyzed using handcrafted features based on potentially imperfect domain knowledge, which while useful may not fully capture all the relevant characteristics inherent in the MEA data. Machine learning, in particular deep learning has the potential to automatically learn relevant characteristics (features) from raw data, without solely relying on handcrafted feature extraction. However, handcrafted features remain critical for encoding domain knowledge and improving model interpretability, especially in scenarios with limited or noisy data, as is often the case in most experimental studies. This study introduces FRAME-C, a knowledge-augmented machine learning pipeline that combines domain knowledge, raw spike waveform data, and deep learning techniques to classify MEA signals and identify ALS-specific phenotypes. FRAME-C leverages deep learning to extract important features from spike waveforms, while also incorporating handcrafted features such as spike amplitude, inter-spike interval, and spike duration, thus preserving key spatial and temporal information. We validated FRAME-C on both simulated and real-world MEA data from human iPSC-derived neuronal cultures, demonstrating its superior performance compared to existing methods for MEA classification.

Figures & tables

Cell line	Sex	Age (years)
sALS 1	M	51
Healthy Control 1	M	50
sALS 2	M	46
Healthy Control 2	M	46
sALS 3	M	50
Healthy Control 3	M	52

Table 1: Donor demographics of the iPSC lines used in this study 0 0 footnotetext: sALS: sporadic amyotrophic lateral sclerosis; M: male; F: female. sALS 2 and Healthy Control 2 are identical twins.

Figure 1: Distribution of mean firing rates (MFRs) of the MEA recordings for each cell line. MFRs of all the MEA signals recorded from human iPSC-derived neuronal cultures, shown per cell line on a log scale. Boxes indicate the interquartile range (25th–75th percentiles), the centre line indicates the median, and whiskers extend to the most extreme values within 1.5 times the interquartile range. Data were obtained from three differentiation replicates per cell line, each with two technical replicates (wells).

Figure 2: Architecture of the proposed FRAME-C pipeline. The raw spike waveforms are augmented using handcrafted features based on domain knowledge. The sequences of spikes are truncated/padded to a fixed length of $len_{spikes}$ , while the sequences of bursts are truncated/padded to a fixed length of $len_{bursts}$ .

Figure 3: UMAP visualization of spike features for sALS 1. The projection is colored by the well number from which the MEA signal was recorded. The emergence of distinct clusters indicates the presence of batch effects within the dataset.

Classification pipeline	Automated feature extraction	Handcrafted features (domain knowledge)	Deep learning classifier
			Informer	TimesNet	CNN	LSTM
Raw MEA (baseline)	Yes	No	49.08	52.43	53.65	49.84
Zhao et al. Zhao et al. (2019)	No	Yes (binary spike locations)	N/A 1 1 1 Not evaluated, since the pipeline by Zhao et al. was originally proposed with a CNN classifier.	N/A 1 1 1 Not evaluated, since the pipeline by Zhao et al. was originally proposed with a CNN classifier.	52.13 ( $-$ 2.83)	N/A 1 1 1 Not evaluated, since the pipeline by Zhao et al. was originally proposed with a CNN classifier.
FRAME-C V1	Yes (from spike waveforms)	No	58.76 (+19.72)	59.49 (+13.47)	57.60 (+7.36)	59.44 (+19.26)
FRAME-C V2	No	Yes	59.77 (+21.78)	61.05 (+16.44)	55.92 (+4.23)	61.17 (+22.74)
FRAME-C V3	Yes (from spike waveforms)	Yes	62.31 (+26.95)	63.49 (+21.09)	62.58 (+16.64)	63.08 (+26.56)

Table 2: Effect of handcrafted features on test accuracy using all eight cell lines 0 0 footnotetext: The effect of using handcrafted features to support automated feature extraction by a deep learning method. To evaluate FRAME-C, experiments were performed with different levels of automated feature extraction and handcrafted features. Accuracies are averaged across all 40 recording days. V1: waveforms only; V2: handcrafted features only; V3: waveforms + handcrafted features. Values are mean test accuracies (%); values in parentheses are the relative change (%) with respect to the raw MEA baseline for the same classifier. Bold values indicate the best result for each classifier.

Classification pipeline	Automated feature extraction	Handcrafted features (domain knowledge)	Deep learning classifier
			Informer	TimesNet	CNN	LSTM
Raw MEA recording	Yes	No	53.35	50.45	53.35	50.60
Pipeline by Zhao et al. Zhao et al. (2019)	No	Yes (binary spike locations)	N/A 1 1 1 Not evaluated, since the pipeline by Zhao et al. was originally proposed with a CNN classifier.	N/A 1 1 1 Not evaluated, since the pipeline by Zhao et al. was originally proposed with a CNN classifier.	56.42 (+5.76)	N/A 1 1 1 Not evaluated, since the pipeline by Zhao et al. was originally proposed with a CNN classifier.
FRAME-C V1	Yes (from spike waveforms)	No	57.71 (+8.17)	62.47 (+23.83)	63.09 (+18.26)	62.29 (+23.10)
FRAME-C V2	No	Yes	64.73 (+21.33)	65.12 (+29.08)	66.21 (+24.11)	66.45 (+31.32)
FRAME-C V3	Yes (from spike waveforms)	Yes	66.87 (+25.35)	68.80 (+36.37)	66.12 (+23.93)	68.14 (+34.62)

Table 3: Test accuracy of FRAME-C after removing unreliable data samples 0 0 footnotetext: Unreliable samples (low-activity cultures, Figure 1 ) were removed before training and evaluation. V1: waveforms only; V2: handcrafted features only; V3: waveforms + handcrafted features. Values are mean test accuracies (%); values in parentheses are the relative change (%) with respect to the raw MEA baseline for the same classifier. Bold values indicate the best result for each classifier.

Classification pipeline	Automated feature extraction	Handcrafted features (domain knowledge)	Deep learning classifier
			Informer	TimesNet	CNN	LSTM
Raw MEA recording	Yes	No	57.92	57.31	53.04	50.00
Pipeline by Zhao et al. Zhao et al. (2019)	No	Yes (binary spike locations)	N/A 1 1 1 Not evaluated, since the pipeline by Zhao et al. was originally proposed with a CNN classifier.	N/A 1 1 1 Not evaluated, since the pipeline by Zhao et al. was originally proposed with a CNN classifier.	75.62 (+42.57)	N/A 1 1 1 Not evaluated, since the pipeline by Zhao et al. was originally proposed with a CNN classifier.
FRAME-C V1	Yes (from spike waveforms)	No	74.78 (+29.11)	72.72 (+26.89)	71.88 (+35.52)	74.05 (+48.10)
FRAME-C V2	No	Yes	64.73 (+11.76)	71.40 (+24.59)	69.81 (+31.62)	71.94 (+43.88)
FRAME-C V3	Yes (from spike waveforms)	Yes	74.25 (+28.19)	73.85 (+28.86)	72.13 (+35.99)	75.08 (+50.16)

Table 4: Test accuracy of FRAME-C for the cell lines derived from identical twins 0 0 footnotetext: Models were trained to distinguish the sALS 2 and Control 2 lines, which were derived from identical twins. V1: waveforms only; V2: handcrafted features only; V3: waveforms + handcrafted features. Values are mean test accuracies (%); values in parentheses are the relative change (%) with respect to the raw MEA baseline for the same classifier. Bold values indicate the best result for each classifier.

Figure 4: Test accuracy of each model across days of neuronal maturation. The change in test accuracy for each model versus the number of days (d) of neuronal maturation post-differentiation.

Feature	Permutation feature importance (%)
Spike duration	23.40
Spike amplitude	1.60
Inter-spike interval	0.00
Burst duration	0.00
Number of spikes per burst	$-$ 0.50
Burst spike rate	$-$ 1.00

Table 5: Importance of each handcrafted feature for classifying simulated MEA recordings 0 0 footnotetext: The permutation feature importance is calculated as the drop in test accuracy when the corresponding feature is removed from the dataset (Eq. 3 ). Bold indicates the most important feature.

Feature	Permutation feature importance (%)
Inter-spike interval	12.30
Spike amplitude	2.90
Number of spikes per burst	1.10
Spike duration	0.80
Burst spike rate	0.40
Burst duration	$-$ 1.00

Table 6: Importance of each handcrafted feature for classifying ALS MEA recordings 0 0 footnotetext: The permutation feature importance is calculated as the drop in test accuracy when the corresponding feature is removed from the dataset (Eq. 3 ). Bold indicates the most important feature.

Figure 5: Longitudinal distribution of handcrafted features and sample yields across ALS and control lines. Boxplots illustrate the variance of extracted features for each donor-derived line across the early (month 1), mid (month 2) and late (month 3) recording timepoints. The accompanying bar chart details the total number of non-zero spike samples recorded per line at each respective stage.

Feature	Timepoint	$N_{\text{Ctrl}}$	$N_{\text{ALS}}$	Median (Ctrl)	Median (ALS)	$p$ -value	Sig.	Cliff’s $\delta$
Peak amplitude ( $\mu$ V)	Early (month 1)	23	23	4.714	4.596	0.1875	ns	$-$ 0.229
	Mid (month 2)	24	24	4.702	4.649	0.3808	ns	$-$ 0.149
	Late (month 3)	24	24	4.669	4.573	0.5848	ns	$-$ 0.094
Spike interval (ms)	Early (month 1)	23	23	643.0	857.5	0.2065	ns	0.219
	Mid (month 2)	24	24	777.0	1059.5	0.1835	ns	0.226
	Late (month 3)	24	24	1370.0	1142.5	0.7650	ns	$-$ 0.052

Table 7: Well-level statistical comparison of handcrafted features between ALS and control lines 0 0 footnotetext: Longitudinal statistical analysis of handcrafted features between ALS and control cell lines aggregated at the well level. No statistically significant differences were observed (

p>0.05

, Mann–Whitney U test) at any timepoint, with small effect sizes (Cliff’s

\delta

) across all metrics. ns: not significant.

Method	No data augmentation (%)	With data augmentation (%)
CNN classifier (Zhao et al. Zhao et al. (2019) )	57.24	56.44
FRAME-C V3 with Informer	63.27	65.92
FRAME-C V3 with TimesNet	65.74	66.76
FRAME-C V3 with CNN	66.28	66.66
FRAME-C V3 with LSTM	65.58	67.82

Table 8: Effect of data augmentation on test accuracy 0 0 footnotetext: Data augmentation is achieved by splitting each recording into multiple segments with overlapping windows. The overlap is configured to produce 10 times more samples compared to non-overlapping splits. Each method is evaluated based on test accuracy. Bold indicates the highest accuracy.

Simulated wells (recordings) per cell type	Pipeline by Zhao et al. (%)	FRAME-C V3 with LSTM (%)
2	51.12	75.06 (+46.83)
4	75.44	82.05 (+8.76)
8	79.13	84.58 (+6.89)
16	94.07	94.21 ( +0.15 )

Table 9: Test accuracy on simulated MEA recordings for different dataset sizes 0 0 footnotetext: Performance comparison of FRAME-C V3 with LSTM and the pipeline by Zhao et al. Zhao et al. (2019) on simulated MEA recordings. Dataset size is measured by the number of simulated wells (recordings) per cell type. Values in parentheses are the relative improvement (%) of FRAME-C over the pipeline by Zhao et al.

Brain Disorder Classification

See all 20 papers

Siam-ALS: An Explainable Siamese Deep Learning Framework for Early ALS Detection Using EEG and Eye-Tracking Signals
Oct 11, 2026Abdur Rahman, A. A. M. Muzahid, Istiak Ahmed +4Clinical Prediction Multimodal Classification
School of Electrical and Electronic Engineering, Shanghai University of Engineering Science, Shanghai, 201620, China. · School of Information Technology, Murdoch University, Murdoch, 6150, WA, Australia. · Department of Information and Communication Engineering, Daffodil International University, Dhaka, 1216, Bangladesh.
Model-agnostic Knowledge Augmented Machine Learning for Weakly Annotated Bioimaging Data
Oct 9, 2026Jayanie Bogahawatte, Sachith Seneviratne, Simon Maksour +6Weakly Supervised Learning
AI, Optimization and Pattern Recognition Research Group, Faculty of Engineering and Information Technology, The University of Melbourne, Melbourne, Victoria, Australia · School of Science and Molecular Horizons, University of Wollongong, Northfields Avenue, New South Wales, Australia · Charles Perkins Centre, The University of Sydney, Sydney, New South Wales, Australia +3
EEG-SpikeAgent: Agentic Closed-Loop Program Synthesis for Automated EEG Spike Detection
Jul 6, 2026Sonali Santhosh, Kelly Shuhong Yu, Eugene Chang +3Electroencephalography LLM-Based Program Synthesis
Dept. of Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, MA, USA · Dept. of Neuroscience, University of California, Berkeley, Berkeley, CA, USA · Dept. of Neurology and Neurologic Sciences, Stanford University, Palo Alto, CA, USA +1

Oct 11, 2026J/K move · Enter open · S save

Siam-ALS: An Explainable Siamese Deep Learning Framework for Early ALS Detection Using EEG and Eye-Tracking Signals

Abdur Rahman, A. A. M. Muzahid, Istiak Ahmed, Dipto Biswas+3

School of Electrical and Electronic Engineering, Shanghai University of Engineering Science, Shanghai, 201620, China. · School of Information Technology, Murdoch University, Murdoch, 6150, WA, Australia. · Department of Information and Communication Engineering, Daffodil International University, Dhaka, 1216, Bangladesh.

Clinical Prediction Multimodal Classification Siamese Neural Networks

Open paper AskPDF arXiv

Amyotrophic lateral sclerosis (ALS) diagnosis remains challenging due to heterogeneous symptoms and reliance on clinical assessment, underscoring the need for objective, data-driven approaches that leverage patient-generated physiological signals. We propose Siam-ALS, an explainable Siamese deep learning framework that jointly models synchronized electroencephalography (EEG) and eye-tracking (ET) signals. EEG captures ALS-related changes in brain activity, while ET provides information about oculomotor control, allowing the model to learn complementary patterns associated with ALS. The framework employs two modality-specific encoders: a convolutional EEG encoder that captures spatiotemporal neural patterns and an MLP-based ET encoder that models oculomotor dynamics. These modality-specific features are processed through weight-sharing Siamese branches and fused into a compact 64-dimensional joint embedding. A lightweight classifier is used to predict ALS, while pairwise metric learning helps the model learn more effective patterns from the small, diverse ALS datasets. Performance was evaluated on the public EEGET-ALS dataset using strict subject-level cross-validation, where Siam-ALS achieved an accuracy of 0.988 and an AUROC of 0.998, demonstrating strong performance compared with classical machine learning, DL, and non-Siamese fusion approaches. Cross-dataset evaluation on the independent BigP3BCI dataset achieved an AUROC of approximately 0.98, supporting the generalization capability of Siam-ALS across different BCI paradigms. To enhance interpretability, six XAI methods were evaluated in the learned embedding space, with SHAP achieving the highest, though statistically comparable, fidelity in identifying important EEG and ET patterns that contribute to ALS predictions. Overall, Siam-ALS provides an accurate, generalizable, and explainable framework for multimodal ALS detection.

Abstract

Brain Disorder Classification

Explore similar work

Siam-ALS: An Explainable Siamese Deep Learning Framework for Early ALS Detection Using EEG and Eye-Tracking Signals

Model-agnostic Knowledge Augmented Machine Learning for Weakly Annotated Bioimaging Data

EEG-SpikeAgent: Agentic Closed-Loop Program Synthesis for Automated EEG Spike Detection