Emotion Detection of People with Autism Ages 10–20: Using Beta-Left and Gamma-Right Temporal

Introduction

AI has recently had a major impact on human medical research and applications that call for sophisticated representation, computationally demanding calculations, and decision-making procedures.^1–4 The primary goal of this work is to identify and detect human emotional responses in individuals with Autism Spectrum Disorder (ASD), which encompasses a wide range of neurodevelopmental conditions. People with ASD frequently display repetitive behavioural patterns, and their emotional responses are strongly associated with difficulties in communication and social interaction.^5,6 Autistic people frequently exhibit irregular emotional behaviour, which can be challenging to understand or identify using standard observational techniques. Self-harm, hand flapping, rocking, and prolonged staring are common behaviours associated with autism. Rather of being misbehaviour, such conduct usually reflects stress, overburden, or unfulfilled needs. Giving physicians and caregivers a novel physiological approach to deduce the internal states of individuals with ASD is one of the goals of our work.^1–4 We investigate an option that uses a simple EEG montage and wearable neurophysiology. Numerous researchers have used EEG signals extensively to analyse human-specific behaviour.^7,8

By developing a distinct differentiator between ASD behavioural response and otherwise normal reactions, this work seeks to build a reference base for autistic behaviour response. The present study creates a testbed for further research in which we assess the EEG signals of people with and without ASD. Keep in mind that the objective is to find tiny, comprehensible signal features that accurately identify symptoms of ASD rather than to construct a model for brain activity in general. In order to facilitate early screening and emotion-aware intervention, such identification is needed.

This method expands on earlier research on emotion recognition utilizing viable, small signal sets, where the researchers looked at fifteen distinct physiological markers. EEG, skin conductivity, blood pressure, heart rate variability (HRV), and other factors were assessed. The importance of EEG signals, which frequently independently disclose respondents' inner emotional status, was one of the study's main conclusions.^1–4

In order to determine if a small collection of EEG characteristics may distinguish ASD from neurotypical controls, this study offers a proof-of-concept analysis using a four-sensor headband. Over a period of ten to fifteen minutes, the EEG bands (alpha, beta, delta, gamma, and theta) are measured. Four distinct sensors are positioned at the AF7/AF8 (frontal) and TP9/TP10 (temporo-parietal) to measure each band.⁹

Five neurotypical controls and seven people with autism spectrum disorder (ASD) are involved in the pilot experiment. The findings set the stage for a longer follow-up with more participants and recordings that are controlled for artifacts.

In summary, the objective of this article aligns with several sustainable development goals (SDGs), such as: SDG 3: Good Health and Well-Being, by helping to identify the emotions of individuals with ASD, which may help clinicians and parents understand their needs. By empowering the educational process for people with ASD through the use of emotion detection tools, it also advances SDG 4: Quality Education. It also aligns with SDG 10: Reduced Inequalities, which is exemplified by reducing educational and social barriers. Additionally, our endeavour advances SDG 9: Industry, Innovation, and Infrastructure by introducing a novel health data-driven innovation.

Methods

Participants and data analysis

The authors of this study examined the EEGs of twelve patients, including five neurotypical controls (n = 5) and seven individuals with ASD (n = 7). Alpha, beta, delta, gamma, and theta bands are the five band-isolated time series that each participant's deployed sensors monitor and report. The data taken by the patients using a written ethical approval received by the ‘Arab Village for Special Challenges’ under approval number (ز/1810). Two sensors were positioned in front of the subject's head for this arrangement, specifically the AF7 sensor on the left and the AF8 sensor on the right, or above the eyes. However, as seen in Figure 1, two additional sensors, TP9 and TP10, were placed toward the rear, or close to the ears (TP9 on the left mastoid region and TP10 on the right mastoid region). The sensor data is reported and stored as channels 1, 2, 3, and 4 using Muse 2 technology. Before statistical testing and modeling, the sessions that some participants completed were combined inside the subject.

Figure 1. Muse sensors setup for EEG measurement.

The most challenging aspect of this study was gathering data from participants with ASD; the authors intend to include at least 30 people with ASD (ages 10–20) from nearby clinics and centers in their next investigation. Participants who can tolerate wearing a head-mounted sensor and have a confirmed diagnosis of ASD will be included. People with other neurological conditions, including epilepsy, that could interfere with EEG recordings will not be included by the authors.^10,11

To put it succinctly, all study methods will be carried out with Institutional Review Board (IRB) permission, and prior to participation, participant assent and parental agreement will be sought.

The Muse 2 headband is used in this study to take measurements.^12,13 Although Muse 2 may record many metrics, the pilot study solely looks at EEG data. The authors of a prior study⁴ employed a NeXus Q32 system with 14 physiological indicators, including EEG, to identify human emotions. EEG signals (alpha, beta, delta, gamma, and theta) may be adequate to identify the emotion without the need to measure additional signals, according to earlier thorough research. Actually, the primary problem with depending solely on EEG is that it is very challenging, if not impossible, to use EEG signals in a non-invasive way.

The measuring sessions take place in a pleasant environment with video observation, and they last only ten to fifteen minutes. Every session involving subjects with ASD was carried out under the guidance and assistance of a therapist, who assisted in keeping the sessions organized and ended them when needed. As a result, this illustrates the challenges the writers encountered when gathering information for their paper.

When using this type of data, where participation is voluntary and withdrawal is possible at any time, confidentiality is obviously essential. An example of the data is shown in Table 1. With an average duration of 15 minutes, each data capture yielded over 250,000 readings, or a reading every 3.6 milliseconds for one normal/healthy (X) and one ASD (Y).

Table 1. Sample data (mean power (μV²) per electrode per band).

Group	Subject	Session	Band	PTP9	PAF7	PAF8	PTP10
Healthy	X_Person_	Session1	alpha	103.3872	48.77104	77.27053	116.2682
Healthy	X_Person_	Session1	beta	82.67007	60.86879	66.44008	102.7144
Healthy	X_Person_	Session1	delta	1825.627	2872.767	954.7735	3646.692
Healthy	X_Person_	Session1	gamma	62.09157	47.28045	35.31806	71.73652
Healthy	X_Person_	Session1	theta	304.5876	182.4597	203.7933	371.7244
Autism	Y_Person	Session1	alpha	3619.203	3450.599	2157.096	2391.242
Autism	Y_Person	Session1	beta	2936.616	3227.962	1660.67	1790.658
Autism	Y_Person	Session1	delta	23297.29	25694.04	18576.49	16746.58
Autism	Y_Person	Session1	gamma	2316.774	2975.258	1704.566	1336.96
Autism	Y_Person	Session1	theta	7521.203	7677.439	4044.538	4844.26

The power per band per electrode (P_band,e) is computed as the mean of squares of the time series. For each subject and band:

$$P_{\text{band},\mathrm{e}}=\frac{1}{N}\sum _{i=1}^N{\left(x_{i,e}\right)}^2$$

Where: P_band,e is the band power at electrode e (TP0, AF7, AF8, TP10); N is the number of reading samples; $x_{i,e}$ is the power at electrode (e) at sample time i. For each subject, 250,000 samples were collected with 5 milliseconds separation between successive readings.

To reduce inter-subject scale differences, the relative power at each electrode is given by

$${\mathit{\text{RelativePower}}}_{\text{band},\mathrm{e}}=\frac{P_{\text{band},\mathrm{e}}}{{\sum }_{\mathrm{b}\in \{\mathrm{\delta },\mathrm{\theta },\mathrm{\alpha },\mathrm{\beta },\mathrm{\gamma }\}}{P}_{\mathrm{b},\mathrm{e}}}$$

Results & Discussion

Figures 2A–2E show the difference in average power alpha, beta, delta, gamma and theta bands; each figure shows the overall power average for the four sensors/channels (AF7/AF8 and TP9/TP10). The blue bars in the figure represent the normal control subjects, while the yellow bars represent the ADS) subjects. The average power of a given band, e.g., alpha, is computed as the average across all participants (e.g. with ADS) over a given sensor, for example TP9. Then we take the average across all ADS participants. Since the sample is small, we use an error range by computing the deviation of the average from the standard deviation. Figures 2A–2E shows the summary of the results, with the yellow bar indicating the relative power average for the ADS sample, while the blue bar represents the normal control sample. The thin black line represents the error bar for each result. For example the Alpha@TP9 autism mean $\approx$ 661.4 with SE (error) ≈ $\mp$ 495.8. In other words, the Alpha@TP9 range is (661.4-495.8) to (661.4+495.8) or minimum of 165.6 to maximum of 1157.2.

Figure 2. Alpha band, Beta band, Gamma band, Delta band, Theta band: Group Means ± SE.

Across all the Figures 2A–2E, the Autism group’s exhibit higher power than the Healthy group, indicating higher absolute power on average. However, the relative β at TP9 and relative γ at TP10, are the clearest discriminators of ASD vs. controls.

The absolute-power differences observed for gamma at TP10 (right temporal) with a Hedges’ g ≈ 0.94 (Autism > Healthy) indicating large effect. For beta at TP9 (left temporal) with g ≈ 0.79 (Autism > Healthy), the difference index is moderate to large.

Using the single–band discrimination based on log10 overall band power (per subject), the gamma band ranked highest (AUC ≈ 0.71), with beta at AUC ≈ 0.69 next highest. Alpha showed modest overall power at (AUC ≈ 0.63), and theta/delta at (AUC ≈ 0.57/0.54).

Figure 3 summarizes the results for the Area under ROC curve, where we utilized the leave-one-subject-out (LOSO) principle. The figure illustrates the Receiver Operating Characteristic (ROC) for the three models (two-feature, log band, and combined 7 features)

Figure 3. Muse sensors setup for EEG measurement.

The two-feature model (Beta and Gamma bands) with AUC = 0.77, reveals a focused classifier using the beta band at TP9 and gamma band at TP10. Note that the entire 2-factor curve sets above the diagonal line, which represents the medium chance (FPR=TPR). The two feature model exhibited the best discrimination between ADS subjects and normal control subjects. Hence, the two features model may be sufficient to do most of the work in distinguishing between the groups. This findings, however, needs to be further investigated with more experiments and subjects, especially in light of other research results, which emphasize the alpha band.²³

The log-bands baseline (AUC = 0.20) is a broader feature set using log-transformed band power across all sensors (alpha, beta, gamma, theta, delta). LOSO. The entire ROC curve falls below the diagonal, which represents the 0.5 chance (FPR=TPR). This indicates weak or inconsistent signal in these generic summaries for this dataset.

The combined model with AUC = 0.71 is a representation of both models, the 2 feature model and the log band model. The combined model outperformed the log-bands baseline. However, it trailed behind the two-feature model. This behavior indicates that adding weak or collinear features may weaken the signal and reduce generalization.

Even, at low false-positive rates, performance remains meaningful despite the small sample. For example, at FPR ≈ 0.20 (1/5 healthy misclassified), the two-feature model detects ≈57% of Autism cases. The combined model (at FPR = 0.2) detects ≈71%. At FPR ≈ 0.40, the same pattern holds (two-feature ≈57%, combined ≈71%).

The alpha effects in our dataset are modest ( Figure 4), although, several studies detect alpha abnormalities in Autism. Figure 4 shows that only at >0.8 FPR, the Alpha band begins to detect autism cases. This does not necessarily contradict the alpha-centric literature, were these studies frequently rely on other sensors such as eyes-closed reactivity, individual alpha peak frequency, or connectivity measures. Our experiments continue to reliably detect autism to a good degree without the use of these extra parameters. Within these constraints, the temporal beta/gamma measures provide strong enough and reliable discrimination between the two sets.

Figure 4. Alpha ROC Plot.

The ROC overlay demonstrates that a minimal, two-feature sensors, namely the relative beta at TP9 and the relative gamma at TP10, deliver the highest cross-validated AUC and favorable sensitivity. This avoids the need to rely on a larger sensor array or complex preprocessing. This simple design is practical for real-world screening. The two temporal sensors are fast to place, and relative features aren’t easily thrown off by changes in overall amplitude. Table 2 provides a summary of key performance/results.

Even with ROC curves and LOSO cross-validation, models can appear overly optimistic especially with relatively low sample. However, in our study this risk is acknowledged, and remedied by aggregating repeated sessions within subject, used leave-one-subject-out (LOSO) to hold out entire participants, and controlled multiple testing via FDR, steps intended to reduce overfitting and selection bias. We also emphasized effect sizes (Hedges’ g) alongside ROC to avoid relying on a single metric and noted that findings are preliminary pending larger, multi-site validation with stricter artifact control and broader coverage, which we outline as future work.