Artificial Intelligence Music Platform for Piano Majors Improves Learning and Reduces Anxiety: Randomized Study with Heart Rate Variability Cues

Summary

This protocol describes a randomized workflow in which an AI-driven music education platform with HRV-informed micro-interventions and visualization-based feedback is used to improve music-theory learning and reduce state anxiety in undergraduate piano majors.

Abstract

Artificial intelligence (AI) is increasingly applied in music pedagogy, yet reproducible workflows that jointly evaluate knowledge acquisition and anxiety regulation remain limited. This protocol describes a randomized controlled workflow for implementing an AI-driven music education platform within routine studio-class teaching and assessing its effects in undergraduate piano majors. Participants are screened and randomly allocated to an AI-driven platform or traditional instruction for 2 weeks (twelve 30-min sessions). Outcomes include a structured music-theory/analysis test and the State-Trait Anxiety Inventory–State (STAI-S; 20 items, 4-point scale; total 20–80). Heart-rate variability (HRV) is recorded during a 5-min seated rest and a 5-min standardized practice segment; primary indices include the root mean square of successive differences of normal-to-normal intervals (RMSSD; short-term vagal activity), the standard deviation of normal-to-normal intervals (SDNN; overall variability), and high-frequency (HF) power (parasympathetic-related spectral component). The protocol enables instructors and researchers to deploy the platform and follow a standardized sequence covering participant screening, randomization execution, platform setup/configuration, HRV sensor placement and recording, artifact handling and HRV computation, administration/scoring of all outcome measures, and final data export, file naming, and secure storage for statistical analysis. Analyses use independent-samples t-tests and analysis of covariance (ANCOVA) with baseline as a covariate. Representative results show higher adjusted post-test theory performance in the AI-driven arm (partial η² = 0.206) and lower post-intervention state anxiety (t(38) = −3.486, p = 0.001; Cohen’s d ≈ 1.10) than controls, with HRV patterns providing physiological context and secondary cognitive-load outcomes indicating reduced extraneous load.

Introduction

In contemporary higher music education, instructional goals extend beyond knowledge transmission to include students’ psychological well-being and sustainable development as performers. Piano majors routinely train under repeated evaluative pressure from juries, competitions, auditions, and public recitals, where anxiety can disrupt attentional control and degrade performance quality¹.

Music performance anxiety (MPA) and examination-related anxiety are consistently reported in conservatory and university samples, and high prevalence estimates have been documented across different cohorts and measurement approaches². Beyond subjective distress, anxiety is associated with physiological arousal, attentional distraction, and reduced self-efficacy, which can translate into pitch/timing instability and inconsistent performance under evaluation³.

Recent advances in artificial intelligence (AI) provide a pathway to more targeted support in this context. AI-driven personalization can adapt content difficulty, pacing, and feedback to learners’ evolving needs and can reduce reliance on one-size-fits-all instruction in studio teaching⁴. In parallel, affective computing can infer momentary emotional states from multimodal cues (e.g., facial expression, speech, and physiological signals), enabling the system to respond to anxiety fluctuations during practice⁵.

A key physiological marker in this domain is heart-rate variability (HRV), which is commonly used as a noninvasive index related to autonomic regulation and stress responsiveness⁶. When embedded in instruction, HRV-informed prompts can support just-in-time micro-interventions (for example, brief paced-breathing or mindfulness cues) that aim to down-regulate acute state anxiety during practice⁷.

A second design pillar is visualization-based feedback. Dashboards that display progress curves, error-location maps, or HRV-informed indicators can strengthen self-monitoring and teacher–student co-regulation by making practice behavior and cognitive–affective load more transparent⁸. From a learning-science perspective, these design choices align with Cognitive Load Theory, which emphasizes reducing extraneous load through streamlined presentation, segmentation, and feedback design while supporting schema construction through germane processing⁹.

Despite rapid development of AI-based music learning tools, relatively few studies in higher music education integrate AI-driven personalization, physiological monitoring (HRV), and visualization-driven feedback within a single platform and evaluate the complete workflow under randomized conditions¹⁰. Moreover, many reports emphasize outcomes without specifying filmable, parameter-level procedures for sensor recording, artifact handling, and data export, limiting reproducibility and cross-site replication. The present work addresses this gap by presenting a step-by-step protocol for deploying an AI-driven music education platform in studio-class routines and assessing its effects on music-theory/analysis achievement and state anxiety (STAI-S), with HRV indices recorded during standardized rest and practice segments to provide physiological context.

Protocol

The study protocol was approved by the Institutional Review Board of the School of Arts, Jingchu Technology University (Approval No.: JCTU-ARTS-2025-013). Written informed consent was obtained from all participants prior to enrollment.

Registration of participants
Obtain written informed consent prior to enrollment after each potential participant receives a written and oral explanation of study objectives, procedures, potential risks and benefits, confidentiality, and data-handling arrangements. Store signed consent forms separately from research data in a locked cabinet or an encrypted digital archive with restricted access limited to the principal investigator and one designated data manager. Record the approval number, approval date, and the consent date for each participant in a study log to support transparent reporting and replication¹¹. Register the study protocol in an appropriate trial registry if required by institutional policy or journal guidance, and record the registration identifier and registration date in the study log¹².

Participant eligibility and recruitment
Screen and recruit participants
Recruit 40 undergraduate piano majors aged 18–22 years from a conservatory or university music program, aiming for an approximately balanced sex ratio with a target deviation of no more than 60:40. Verify that each candidate has completed at least 6 months of formal piano study, can read staff notation fluently, and has introductory music-theory knowledge sufficient to complete the baseline test. Invite interested candidates to a screening visit and provide written study information describing time commitments, data collection procedures, and confidentiality protections at least 24 h before screening.

Apply inclusion criteria
Administer the Generalized Anxiety Disorder-7 (GAD-7) questionnaire to each candidate and score it according to the instrument manual¹³. Conduct a brief semi-structured interview lasting 5–8 min to confirm that the candidate can understand instructions, tolerate procedures, and attend all sessions. Include participants only if all criteria are met: Age 18–22 years, at least 6 months of formal piano training, GAD-7 score between 5 and 14, indicating mild to moderate anxiety¹⁴, Willingness to complete all intervention and assessment procedures, and provision of written informed consent.

Apply exclusion criteria
Exclude candidates with current severe mental disorders or ongoing psychiatric treatment that could interfere with participation or outcome validity. Exclude candidates who have participated in similar AI-driven music-training studies within the previous three months, defined as any intervention study involving algorithm-driven practice guidance or automated feedback for ≥3 sessions. Exclude candidates who are unfamiliar with the target repertoire and cannot achieve a stable baseline performance after a standardized familiarization period of 15 min with the excerpt. Exclude candidates with major motor, visual, or hearing impairments that would interfere with piano performance or heart-rate variability (HRV) recording. Consider excluding candidates with conditions or factors that substantially compromise HRV measurement quality (e.g., known arrhythmias, implanted pacemakers, or medications that markedly affect autonomic function), and document the rationale in the screening log when applicable¹⁵. Document all inclusion and exclusion decisions, including reasons for exclusion, in a screening log identified by anonymized participant IDs, and store the log separately from outcome data.

Randomization and allocation concealment
Preparation of randomization sequence
Use statistical software to generate a 1:1 permuted-block randomization sequence assigning participants to the experimental arm (AI-driven platform) or the control arm (traditional instruction). Predefine the block size used for permuted blocks as 4, and record the software name, version, and a fixed random seed value of 20250113 in a randomization log to ensure reproducibility¹⁶. Ensure that the researcher generating the sequence is not involved in participant assessment, outcome scoring, or teaching. Print individual group assignments on identical paper slips (approximately 5 cm x 2 cm) and place each assignment into a sequentially numbered, opaque, sealed envelope. Record envelope preparation details in an audit log (date, preparer, total envelopes = 40) and store the audit log separately from outcome data.

Implementation of allocation concealment
Store sealed envelopes in a secure location accessible only to personnel responsible for allocation. At the end of each participant’s baseline session, retrieve the next envelope in numerical order. Open the envelope in the presence of the participant and record the revealed allocation in an allocation log (participant ID, envelope number, date/time, allocator signature). Keep all remaining envelopes sealed until use to preserve allocation concealment.

Maintaining blinding of performance raters
Assign each participant an anonymized ID that does not encode group allocation using the format P001–P040. Label all audio and video recordings with participant ID and timepoint (pre- or post-intervention) only. Provide performance raters with a de-identified, counterbalanced set of recordings containing no information about group allocation and with file metadata removed prior to sharing.

Repertoire and setting
Selection of the target repertoire
Use Mozart’s Piano Sonata in G major, K.283, Movement II (Andante) as the common training and assessment piece. Use bars 1–32 as the standardized excerpt for training and assessment, and confirm that this excerpt length and difficulty are appropriate for intermediate-level piano majors, allowing stable tempo targets and consistent phrasing demands across sessions¹⁷.

Standardization of the physical setting
Conduct all sessions in a quiet studio with stable lighting and temperature maintained at 22 ± 1 °C. Position a weighted-key digital piano or an acoustic piano with comparable key action in a fixed room location and keep the bench height constant for each participant across visits. Place a microphone 0.8 m from the instrument at a fixed height of 1.2 m, angled 45° toward the soundboard or keybed to capture a clear signal without clipping. Mount a camera on a tripod 1.8 m from the performer at a height of 1.4 m, capturing hands and upper body at a consistent angle across all sessions.

Calibration of equipment before each testing day
Set and test the metronome volume so that it is clearly audible to the participant but does not dominate the audio recording, targeting a playback level of approximately 60–65 dB at the bench position. Adjust recording gain on the audio interface or recording device to avoid clipping while preserving a high signal-to-noise ratio, targeting peak levels between −12 dB and −6 dB during forte passages. Check camera framing and focus to ensure that the keyboard, hands, and upper torso are consistently in view. Verify HRV device connectivity and signal quality by observing the real-time display and confirming stable R–R interval detection (or stable ECG waveform) for at least 30 s before starting data collection. If audio clipping occurs, reduce input gain by 3–6 dB, repeat a 10 s test recording, and proceed only after peak levels remain below the clipping threshold. If R–R detection is unstable, increase skin contact by tightening the chest strap and moistening contact pads (or reposition ECG electrodes per manufacturer guidance), repeat the 30 s signal check, and proceed only after a stable trace is obtained.

Outcome measures
Knowledge and analysis (primary cognitive outcome)
Develop two parallel forms (Form A and Form B) of a structured music-theory and piece-analysis test matched on content domains and difficulty¹⁸. Include 20 items covering harmony (8 items), rhythm segmentation (4 items), motivic recognition (4 items), and formal analysis (4 items) of the target repertoire and related materials. At each assessment timepoint, seat the participant alone in a quiet room and provide either Form A (baseline) or Form B (post-test), according to a predefined schedule. Instruct the participant to complete the test within 30 min without using instruments, textbooks, or electronic aids. Collect the response sheet at the end of the test and record completion time in minutes. Score each response using a standardized answer key and calculate the total accuracy score (0–100) as the primary cognitive outcome, awarding 5 points per item. Store the total score and completion time under the participant’s anonymized ID.

Performance rubric (secondary skill outcome)
Record an audio–video performance of the target excerpt for each participant at both pre- and post-intervention using the standardized instrument, microphone placement, and camera angle. Use a performance rubric with four dimensions: melodic fluency, rhythmic accuracy, coordination and technique, and expressive interpretation. Rate each dimension on a 1–5 scale to yield a total score between 4 and 20. Train at least two independent raters using a detailed rubric document and a set of 6 pilot recordings (3 baseline and 3 post-test recordings from non-study students) until acceptable agreement is achieved on all dimensions. Provide raters with the full set of recordings labeled only by anonymized ID and timepoint, without group information. Instruct raters to evaluate recordings independently and to record their scores in a secure scoring sheet or database. Identify any discrepancies greater than 2 points on a single dimension and resolve them through discussion; if consensus cannot be reached, obtain a third rater’s judgment. Before full scoring, estimate inter-rater reliability using the intraclass correlation coefficient (ICC) on the 6 pilot recordings and confirm that ICC is at least 0.80¹⁹. Summarize group-level rubric scores (total and four dimensions) at pre-test and post-test in Supplementary Table 1, and report inter-rater reliability statistics (ICC with 95% confidence intervals) in Supplementary Table 2.

State anxiety and heart-rate variability (secondary affective/physiological outcomes)
Administer the State-Trait Anxiety Inventory–State (STAI-S; 20 items, 4-point Likert scale, total score 20–80) at both pre- and post-intervention²⁰. Instruct participants to respond based on how you feel right now and to answer all items. Check each questionnaire for missing responses and, if needed, ask the participant to complete any omitted items before leaving. Calculate the total STAI-S score according to the instrument manual. If a validated 10-item subset or adapted format is used for visualization or exploratory analysis, document the adaptation and selection rule, provide justification and references, and report Cronbach’s α for both the full scale and the subset in the results²¹. Instruct participants to refrain from vigorous exercise, caffeine, and heavy meals for at least 2 h before HRV recording and to avoid alcohol for at least 12 h before each visit; confirm compliance verbally and record compliance as yes/no in the session log²². Record HRV using either (i) a chest-strap ECG sensor or (ii) a three-lead ECG system. Use ECG-derived R–R intervals with a sampling frequency of 1000 Hz for all recordings, and record the modality (ECG), sampling frequency (1000 Hz), and recording date/time in the session log. For chest-strap recording, position the strap at the level of the xiphoid process with firm skin contact; for ECG recording, place electrodes in a standard lead II configuration and verify stable R-wave detection before starting. For resting HRV, seat the participant comfortably with feet flat on the floor and hands resting on the thighs; instruct quiet wakefulness with minimal movement and no speaking for 5 min; start recording after a 60 s stabilization period and analyze the subsequent 300 s segment.

For practice HRV, instruct the participant to perform the target passage for 5 min at a fixed tempo of 76 beats per minute and a fixed dynamic target of mezzo-forte; use a metronome to maintain tempo and begin analysis after the first 30 s, analyzing the subsequent 270 s segment. After each recording, export the raw R–R interval series as a comma-separated values file (CSV) containing at minimum two fields (timestamp in ms; R–R interval in ms), and inspect the trace using the recording software and visual inspection to identify signal loss, artifacts, and ectopic beats.

Apply a prespecified artifact correction workflow and record the proportion of corrected beats for each segment in the session log²³.
Artifact detection rule: flag an R–R interval as an artifact if it deviates by more than 20% from the median of the preceding 5 normal intervals, or if absolute R–R values fall outside 300–2000 ms. Correction method: replace flagged intervals using cubic-spline interpolation of adjacent normal intervals after removing ectopic beats.
Quality-control threshold: repeat the recording segment if the corrected-beat proportion exceeds 5.0% or if there is any continuous signal loss longer than 5 s.

Compute time-domain indices including RMSSD and SDNN using analysis software version 1.3.0, and store outputs in a structured results file (CSV) linked to participant ID, timepoint, and segment type (rest or practice). Compute frequency-domain indices including HF power in the 0.15–0.40 Hz band using a Welch periodogram with a 120-s window length, 50% overlap, Hamming window, and 4 Hz resampling of the interpolated R–R series; report HF in ms², and log-transform HF prior to inferential analysis. Optionally compute LF power (0.04–0.15 Hz) and the LF/HF ratio and report these descriptively with appropriate caution²⁴.

Cognitive load (exploratory secondary outcome)
Administer the multidimensional cognitive load questionnaire developed by Leppink and colleagues at post-test²⁵. Present the 10 items forming three subscales: extraneous load (3 items), intrinsic load (3 items), and germane load (4 items), each rated on a 7-point Likert scale from 1 (strongly disagree) to 7 (strongly agree). For each subscale, calculate the mean of corresponding items to yield a subscale score between 1 and 7; treat missing values as invalid and require complete responses for scoring. Interpret higher extraneous load as greater irrelevant processing, higher intrinsic load as greater inherent task complexity, and higher germane load as greater productive, schema-building effort. When using a non-English version, perform translation and back-translation by two independent bilingual experts, pilot-test the translated instrument in 10 non-study piano majors, and compute Cronbach’s α for each subscale; adopt α ≥ 0.70 as the adequacy criterion²⁶. Report subscale-specific α values in the Results and analyze the three subscales separately, applying Holm–Bonferroni adjustment for multiple comparisons with a familywise α = 0.05²⁷. If desired, derive a composite cognitive-load index as the arithmetic mean of all 10 items.

Baseline (pre-test; Week 1)
Confirm eligibility and informed consent at the beginning of the baseline session and assign the participant an anonymized ID. Record demographic information and training history, including age, sex, years of piano study, and typical weekly practice hours, in a baseline data form. Administer the knowledge and analysis test (Form A) in a quiet environment, enforce the 30 min time limit, and record the total score and completion time under the participant’s ID. Attach the HRV sensor and record a 5 min resting HRV segment, followed by a 5 min standardized practice HRV segment of the target excerpt at 76 beats per minute and mezzo-forte dynamic target.

Monitor signal quality during acquisition and repeat any segment with excessive artifacts after adjusting strap tension or electrode placement; repeat when the corrected-beat proportion exceeds 5.0% or when continuous signal loss exceeds 5 s, and document repetition and the reason in the session log. Summarize HRV data-quality indicators (artifact-corrected beat proportion, repeated segments, and exclusions) in Supplementary Table 3. Administer the STAI-S and calculate the total score (20–80) according to the scoring key. Record a baseline audio–video performance of the target excerpt using the predefined instrument, microphone, and camera setup, ensuring consistent framing and audio levels. Label all data files (tests, HRV recordings, questionnaires, and performance videos) with anonymized participant ID and the pre timepoint indicator, and store them in a secure, access-controlled, backed-up directory using the following file naming convention: P###_TP(pre/post)_MOD(test/HRVrest/HRVprac/STAI/perf)_YYYYMMDD_HHMM²⁸.

Intervention sessions (Weeks 2–3)
Scheduling
Schedule 12 supervised sessions of 30 min for each participant in the experimental arm over a 2 week period. Schedule sessions at relatively consistent times of day for each participant within a fixed window of ±2 h relative to that participant’s first-session start time, and record start time and end time for every session to reduce and quantify diurnal variation in HRV and anxiety. Record attendance, reasons for missed/rescheduled sessions, and any protocol deviations in an adherence log. Standardize pre-session conditions by confirming that participants complied with the pre-recording restrictions for caffeine, vigorous exercise, heavy meals, and alcohol, and record compliance as yes/no with brief notes when no²⁹.

Experimental arm: AI-assisted platform (Figure 1)
At the start of each session, instruct the participant to log into the AI platform using the anonymized participant ID, and confirm that no personally identifiable information is displayed in the interface. Load the Mozart sonata excerpt in the interactive score interface and verify that the correct piece, excerpt boundaries (bars 1–32), and difficulty settings are displayed. Configure the platform to record practice telemetry at the note and bar level, including pitch accuracy (%), onset timing deviation (ms), tempo drift relative to metronome (%), bar-level error locations, and repetition counts for each looped segment; set the telemetry update interval to 200 ms and the bar-level aggregation interval to 1 bar. Run a 45-s calibration playthrough of the excerpt to confirm that the platform captures note events and aligns detected notes to score positions; repeat calibration if alignment errors exceed either of the following tolerances: median onset mismatch greater than 80 ms over any 10-note window or repeated bar-level misalignment across 2 consecutive bars.

Enable the personalized training module so that it ingests within-session performance features and produces segment-level goals using a fixed rule: select the two bars with the highest combined error score, where combined error score = 0.5 x pitch error rate + 0.3 x absolute timing deviation + 0.2 x tempo drift; record selected bars and the top three error drivers in the session log. Instruct the participant to practice targeted segments while the interactive score highlights notes in real time, supports looping of difficult passages, and plays MIDI exemplars of target segments at the target tempo (76 bpm) and at a reduced tempo of 60 bpm for scaffolding. Allocate session time using a fixed structure and record realized minutes for each phase: 3 min login/calibration, 21 min targeted segment practice, 5 min reflection with the conversational partner (Section 7.3), and 1 min wrap-up/export and notes³⁰.

Conversational practice partner
Activate the conversational practice partner powered by a large language model (LLM) through a secure API connection, and record the model version as LLM-PracticePartner v2.1 (build 2025-01-15), the inference endpoint type as cloud, and the access-control method as token-based authentication in the study log. Configure the partner with a fixed system prompt that constrains outputs to practice-related content (phrasing, tone-color intentions, pedaling decisions, error diagnosis, and next-step suggestions), and disable any functions that request personal data.

After each targeted practice segment, prompt the participant to describe phrasing choices, tone-color intentions, pedaling decisions, perceived difficulties, and perceived anxiety level on a 0–10 self-report scale, and record the responses as structured fields in the session log. Configure the partner to return short segment-specific suggestions aligned with current goals using fixed generation parameters: maximum response length 80 words, maximum 2 turns per segment, temperature 0.2, and a 2-s response timeout; record the number of exchanges and total dialogue duration per session. Ensure that dialogue time remains secondary to playing by capping total dialogue time per session at 6 min and enforcing the cap by disabling further prompts once the cap is reached.

Privacy and security: store conversation logs using anonymized IDs, encrypt in transit using TLS 1.2 or later and at rest using AES-256, restrict access to authorized personnel only, set raw transcript retention to 24 months, and record the deletion date for each participant’s transcript set in the data-management plan³¹.

Multimodal emotion recognition and micro-interventions
Enable the multimodal emotion-recognition module and record the input modalities and sampling/update rates in the session log as follows: facial video at 30 frames per second, voice audio at 16 kHz mono, and HRV stream updated at 1 Hz using ECG-derived R–R intervals. Define elevated stress using prespecified trigger logic and record the exact parameters.
HRV trigger: RMSSD falls at least 1 within-participant baseline standard deviation below the baseline resting RMSSD and remains below this threshold for at least 60 consecutive seconds during practice, calculated on a rolling 30 s window updated every 1 s.
Performance trigger: running error rate exceeds the participant’s mean error rate for the previous three sessions by at least 20% for a continuous window of at least 60 s, where running error rate is computed over a rolling 20 note window updated every 5 notes.
Multimodal confirmation rule: trigger requires either the HRV trigger alone or the performance trigger plus at least one corroborating cue from face/voice features, defined as either facial valence score decreasing by at least 0.30 from the participant’s within-session baseline for at least 10 consecutive seconds or voice jitter increasing by at least 15% relative to the participant’s within-session baseline for at least 10 consecutive seconds, with baselines computed from the first 2 min of the session.

Specify whether multimodal fusion is rule-based thresholding or a trained classifier. Use rule-based thresholding for this protocol and record all thresholds in the platform configuration file³². Program the system to pause practice when the stress threshold is met and to trigger one micro-intervention selected from a prespecified library. Record intervention type, start time, duration, and completion status in the session log.
Paced breathing: 90 s script at a fixed breathing cadence of 6 breaths per minute delivered by on-screen timer and audio cue³³.
Brief mindfulness cue: 75 s guided attention instruction with fixed wording, presented as on-screen text with audio narration.
Low-arousal background music: 90 s playback of a prespecified low-arousal track at a fixed loudness target of 55–60 dB at the bench position, with a track identifier recorded as Track-LA-01 in the session log.

After the micro-intervention, resume the interrupted passage and reduce tempo by 8% for the first 60 s of re-engagement when the trigger was HRV-based, then return to target tempo if running error rate returns to within 10% of the participant’s mean error rate for the previous three sessions for at least 30 consecutive seconds.
Quality-control: record the number of triggers, the number of completed micro-interventions, and the number of aborted interventions per session, and compute a compliance ratio for each participant as completed triggers divided by total triggers.

Validate trigger performance by summarizing false-positive and false-negative events using prespecified criteria.
False positive: a trigger occurs when the participant reports low anxiety (≤2/10) at the next prompt and running error rate remains within 10% of the participant’s mean error rate for the previous three sessions for the subsequent 60 s.
False negative: participant reports high anxiety (≥7/10) at the next prompt during a segment without a trigger despite running error rate exceeding the participant’s mean error rate for the previous three sessions by at least 20% for at least 60 s or despite HRV meeting the HRV trigger criterion.

Summarize false-positive and false-negative counts and rates per participant and per session in Supplementary Table 4³⁴. Report the composite cognitive-load index descriptively in Supplementary Table 4.

Logging and adherence
Configure the platform to log all practice events and outputs as structured records linked to anonymized ID, session number, and timestamp in ISO 8601 format. Required fields include targeted bars, loop counts, accuracy metrics, tempo drift, dialogue counts and total dialogue duration in seconds, trigger events, intervention type and duration in seconds, and any manual overrides by instructors. Export session-level summaries at 20:00 each day, and store both raw logs and summarized tables in a secure server with automated backups scheduled once daily at 02:00. Use a standardized file naming convention for all exports formatted as StudyID_Group_ID_Session##_YYYYMMDD_HHMM and save in non-proprietary formats as follows: CSV for logs, WAV for audio recordings, MP4 for video recordings, and PDF for session reports. Record the export path, file checksum using SHA-256, and backup confirmation in the study log to support traceability³⁵.

Control arm: traditional instruction
Schedule twelve 30 min teacher-guided sessions over 2 weeks for each participant in the Control arm, matching total supervised contact time of the Experimental arm³⁶. Use the same repertoire, teaching content, and instructors as in the Experimental arm to control teacher and material effects³⁷. Conduct each session using standard studio pedagogy, including teacher demonstration, verbal explanation, and targeted exercises on the excerpt. Assign home practice tasks using conventional verbal or written instructions without AI-generated prompts or analytics, and define the home task as one 20 min self-practice block per day focused on bars 1–32. Withhold individualized analytics, automated prompts, dashboards, and micro-interventions from Control participants. Provide each Control participant with a brief paper or electronic practice log and instruct recording of date and approximate duration of additional unsupervised practice in minutes. Collect logs at the end of the intervention and enter total unsupervised practice time for the target excerpt into the study database.

Visual feedback and self-regulation pathway (Experimental arm)
Enable a graphical feedback system linking practice behavior, emotional fluctuations, and physiological load on a shared time axis³⁸. Compute emotional heat maps based on multimodal inputs, including voice characteristics, HRV trends, and anomalies in performance behavior such as error spikes or tempo instability³⁹. Configure daily summaries to flag hot spots defined as any 30 s interval in which error rate is within the top 20% of that session and at least one stress marker is present (HRV trigger met or multimodal confirmation met), and record flagged hot spots for subsequent sessions⁴⁰. Generate HRV trend charts summarizing RMSSD and HF power for three fixed epochs per session: pre-practice (first 60 s), mid-practice (minutes 3–4), and post-practice (final 60 s), using consistent units and axis scaling across participants⁴¹. Identify marked drops in RMSSD or HF power defined as a decline of at least 15% relative to the session pre-practice epoch aligned with a flagged hot spot and treat these patterns as candidates for platform-guided interventions; record whether an intervention was triggered and completed⁴².

Plot progress curves for pitch accuracy, rhythmic accuracy, musical completeness, and error rates across sessions using session-level means and 95% confidence bands computed across participants within each group⁴³. Overlay performance-progress curves with the emotional heat map to support reflection on emotion–performance coupling and metacognitive control⁴⁴. Provide instructor dashboards displaying the same information at individual and group levels, updating dashboards at the end of each session within 5 min of session completion and highlighting peaks in anxiety or performance instability⁴⁵. Use dashboards to inform tailored pedagogical responses documented in the instructor notes, including temporary tempo reduction, graded exposure to difficult bars, additional verbal encouragement, or brief one-on-one coaching focused on identified hot spots⁴⁶.

Post-intervention session (post-test; end of Week 3)
Schedule the post-intervention assessment at the end of Week 3 and match baseline conditions as closely as possible, including room, equipment, and time of day within ±2 h of the baseline session start time. Attach the HRV sensor and record a 5 min resting HRV segment followed by a 5 min standardized practice HRV segment using the same placement, sampling settings, and artifact-handling rules as at baseline, and use the same tempo target of 76 beats per minute and mezzo-forte dynamic target⁴⁷.

Administer the knowledge and analysis test (Form B) under the same conditions and time limit used at baseline and record total score and completion time⁴⁸. Administer the STAI-S and calculate the total score (20–80) according to the scoring key⁴⁹. Record a post-intervention audio–video performance of the excerpt using the same microphone and camera setup, instrument, and framing as at baseline⁵⁰. Administer the cognitive load questionnaire and record subscale scores for extraneous, intrinsic, and germane load⁵¹.

For the Experimental arm, conduct a brief semi-structured interview lasting 6–8 min on platform usability and perceived effects on anxiety regulation and learning using a fixed interview guide and consistent prompting⁵². Audio-record interviews with consent and transcribe verbatim for qualitative analysis.

Statistical analysis
Perform an a priori power analysis before recruitment using GPower (version 3.1 or later) and record the test family as F tests and the statistical test as ANCOVA: Fixed effects, main effects and interactions. Specify the ANCOVA model with two groups, baseline score as a covariate, α = 0.05, desired power = 0.80, and assume an effect size of f = 0.30 for the primary cognitive outcome based on prior literature and feasibility constraints. If the assumed effect is expressed as partial η², convert to Cohen’s f using the prespecified conversion f = √(η² / (1 − η²)) and record the calculation method⁵³.

Document the resulting required sample size and confirm whether n = 40 meets or approaches the target under the stated assumptions; retain the power-analysis output file in the study records. Conduct analyses in SPSS 26.0 or equivalent software and record software name/version in the analysis log. Summarize continuous outcomes as mean ± standard deviation and categorical variables as counts and percentages.

Evaluate primary efficacy using between-group comparisons of post-test outcomes with baseline as a covariate (ANCOVA) and report adjusted means, partial η², and 95% confidence intervals⁵⁴. Use independent-samples t-tests for between-group comparisons when appropriate and paired t-tests for within-group changes; apply non-parametric alternatives when assumptions are violated. Check normality and homoscedasticity; log-transform skewed spectral HRV variables before analysis and report transformation details.

When robust methods are used, specify the exact method and rationale. Report two-tailed p-values with α = 0.05 and accompany hypothesis tests with effect sizes (Cohen’s d, partial η²) and 95% confidence intervals. Conduct exploratory analyses in the Experimental arm examining associations between platform usage metrics (total platform minutes, micro-intervention counts, adherence ratio) and cognitive, affective, and HRV outcomes.

Data quality and safety
Inspect physiological signals during each HRV recording for artifacts or unstable R–R detection using real-time display. Repeat any HRV segment with poor signal quality after adjusting strap tension or electrode placement, and repeat when corrected-beat proportion exceeds 5.0% or continuous signal loss exceeds 5 s; record repetition and suspected cause in the session log. Calibrate audio levels and metronome settings at the beginning of each visit to avoid clipping and maintain comparable loudness across recordings, targeting peak levels between −12 dB and −6 dB.

Monitor participants for adverse events, excessive fatigue, or discomfort related to practice or sensors and document events in a safety log. Allow additional breaks or shorten a session if a participant reports undue fatigue, anxiety, or discomfort, and record modifications as protocol deviations. Log all protocol deviations (missed sessions, incomplete HRV recordings, major departures) and consider deviations in sensitivity analyses.

Data handling and blinding
Assign each participant a unique anonymized ID and use this ID to label all recordings, questionnaires, and data files. Store all data on access-controlled drives with regular automated backups scheduled once daily at 02:00 according to institutional data-protection policies. Maintain a separate linkage file mapping participant IDs to personal identifiers in an encrypted location accessible only to authorized personnel. Keep performance raters blinded by providing anonymized and time-shuffled audio–video clips and withholding group allocation information.

Finalize the scoring rubric before rating begins and provide written instructions and exemplar recordings⁵⁵. Recheck inter-rater reliability on a pilot subset of 6 recordings before full scoring and confirm ICC ≥ 0.80.

Ensure that at minimum the following functional specifications are reported: HRV sampling frequency = 1000 Hz, facial video capture = 30 frames per second at 1920 × 1080 resolution, voice audio capture = 16 kHz mono, and platform telemetry update interval = 200 ms.

Timing and adherence
Define the per-protocol adherence threshold as completion of at least ten of the twelve supervised sessions over Weeks 2–3, with each completed session defined as a supervised session lasting at least 27 min of active participation within the planned 30 min slot. Record the number of completed supervised sessions and total supervised minutes for each participant in an adherence log, and compute supervised-minute adherence as total supervised minutes divided by 360 planned minutes expressed as a percentage. Record self-reported additional practice time from participant logs and enter totals into the database as minutes per day and total minutes over the 2 week intervention period, and flag any daily value exceeding 120 min for verification.

For the Experimental arm, record the number, type, start time, and duration of system-triggered micro-interventions per session from platform logs, and compute an intervention exposure index as the total intervention duration in seconds per session. Include all randomized participants in intention-to-treat analyses, and treat missing post-test outcomes using last observation carried forward for questionnaire outcomes and complete-case analysis for performance recordings when re-recording is not possible; record the handling rule in the analysis log. Conduct sensitivity analyses restricted to participants meeting per-protocol threshold and additionally conduct a sensitivity analysis excluding sessions with protocol deviations classified as major, defined as either a missed HRV recording segment or a session shortened by more than 5 min.

Results

Effects on music-theory cognition
As shown in Table 1, the Experimental group (M = 72.45, SD = 6.83, n = 20) and the Control group (M = 71.30, SD = 7.14, n = 20) did not differ at pre-test, t(38) = 0.625, p = 0.604, indicating baseline comparability for the primary cognitive endpoint. After the intervention, ANCOVA controlling for pre-test scores showed higher adjusted post-test means in the Experimental group than in the Control group (adjusted M = 84.95 vs. 78.75; Table 2

Discussion

This randomized study evaluated an AI-driven music-education platform that integrates personalized practice analytics, multimodal emotion recognition with just-in-time micro-interventions, and visualization-based feedback. Relative to traditional instruction, the Experimental arm showed a moderate-to-large gain in music-theory and analysis performance (partial η² = 0.206) and a large reduction in state anxiety (t = −3.486, p = 0.001; d ≈ 1.10), accompanied by a more favorable cognitive-load pr...

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
AI-driven music-education platform (research build)	In-house / Institutional	Platform build ID: v2.1 (build 2025-01-15)	Interactive score + practice telemetry + dashboards + rule-based trigger engine; telemetry update interval = 200 ms; export logs as CSV.
Alcohol wipes (skin prep)	NA (any medical supplier)	NA	For improving electrode/strap contact and signal stability; record lot if available.
Chest-strap ECG HRV sensor (1000 Hz)	Polar Electro	Polar H10; Part No. 92075957	ECG-derived R–R intervals; sampling frequency = 1000 Hz (per your protocol); used for 5-min rest + 5-min standardized practice segments.
Cognitive Load Questionnaire (Leppink, 10 items)	NA (publication source)	NA	7-point Likert; compute 3 subscales; if translated: translation–back-translation + pilot n=10; report Cronbach’s α by subscale.
Computer for platform operation and logging	NA	NA	Any modern laptop/desktop; record OS version; ensure time sync for log timestamps.
Data storage (encrypted drive / secure server)	NA	NA	Access-controlled; automated backups; file naming: “P###_TP(pre/post)_MOD(…)_YYYYMMDD_HHMM”; store raw + derived outputs.
ECG electrode gel / conductive gel (optional)	NA (any medical supplier)	NA	Use if signal quality is unstable; document usage yes/no in session log.
External microphone (if using mic + interface instead of recorder)	Audio-Technica	AT2020	Optional alternative to portable recorder; keep mic distance 0.8 m; avoid clipping; document gain settings.
HRV processing software	Kubios (or equivalent)	NA	Must support artifact detection + interpolation, RMSSD/SDNN, Welch HF (0.15–0.40 Hz) with 120-s window, 50% overlap, Hamming window, 4 Hz resampling; record software name/version used.
Metronome	Korg	MA-2 BLBK	Used to standardize practice tempo (76 bpm) and support consistent timing across sessions.
Portable audio recorder (voice capture)	Zoom	H1n	Record voice audio capture at 16 kHz mono (downsample if recorded higher); store as WAV; keep mic placement constant.
Power analysis software	G*Power	3.1 (or later)	Used for a priori power analysis; record test family/model settings in the power-analysis output file.
Practice-partner large language model (LLM) service	Institutional / Secure API provider	LLM-PracticePartner v2.1 (build 2025-01-15)	Endpoint type: cloud; token-based authentication; store only anonymized IDs; log prompts/outputs as encrypted text.
Questionnaire: GAD-7	NA (licensed/publication source)	NA	Use validated language version; scoring per manual; keep copies in study binder; store responses by anonymized ID.
Questionnaire: STAI-S (20 items)	NA (licensed/publisher)	NA	Use official/validated language version; total 20–80; if showing 10-item subset for figures, document selection rule + reliability in Results.
Statistical analysis software	IBM	SPSS Statistics 26.0	Used for ANCOVA and t-tests; record exact version/build in analysis log.
Tripod / camera mount	Manfrotto (or equivalent)	NA	Any stable tripod acceptable; fix distance/height across sessions (e.g., 1.8 m distance, 1.4 m height). If you have a confirmed model, replace NA with your lab’s model number.
USB audio interface (if using mic + interface)	Focusrite	Scarlett 2i2 (3rd Gen)	Optional; target recording peaks between −12 dB and −6 dB during forte passages (per your protocol text).
Video camera (1080p, 30 fps)	Logitech	C920; Part No. 960-001055	Facial video capture = 1920×1080, 30 fps (as specified in protocol); ensure no identifiable info displayed.
Weighted-key digital piano (or acoustic piano)	NA	NA	88-key weighted action recommended; keep same instrument for all sessions; document instrument type and key action.