Higher-order sonification of the human brain Authors: Francisco-Shu Kitaura, Emi-Pauline Kitaura, Niels Janssen, Antonella Maselli, Ernesto Pereda, Aurelio Carnero Rosell Published: 2025 ABSTRACT AbstractSonification, the process of translating data into sound, has recently gained traction as a tool for both disseminating scientific findings and enabling visually impaired individuals to analyze data. Despite its potential, most current sonification methods remain limited to one-dimensional data, primarily due to the absence of practical, quantitative, and robust techniques for handling multi-dimensional datasets. We analyze structural magnetic resonance imaging (MRI) data of the human brain by integrating two- and three-point statistical measures in Fourier space: the power spectrum and bispectrum. These quantify the spatial correlations of three-dimensional voxel intensity distributions, yielding reduced bispectra that capture higher-order interactions. To illustrate the potential of the approach, we focus on one representative reduced-bispectrum configuration (Q019036) in the OASIS-3 dataset (864 imaging sessions) and provide audio renderings for five age groups (40–50, 50–60, 60–70, 70–80, 80–100 years). As context, prior work with this configuration reported an age-prediction MAE of 4.2 years based on neural networks, trained on bispectrum-derived features–not through sonification itself. But here we emphasize sonification rather than predictive benchmarking. Finally, we treat these audio examples as exploratory illustrations rather than a perceptual evaluation. Our results demonstrate that the information loss (e.g., normalized mean squared error) during the reconstruction of the original bispectra, specifically in configurations sensitive to brain aging, from the sonified signal is minimal. Future studies should include systematic statistical inference and perceptual validation to assess the robustness and perceptual utility of the method. Nevertheless, the approach presented here already provides a general framework for encoding multi-dimensional data into time-series-like arrays suitable for sonification, thereby opening new avenues for scientific exploration and improving accessibility for a broader audience.Subject terms: Neurodegenerative diseases, Neural ageing, Perception The technique of sonification has emerged as a valuable tool in various fields for its ability to transform complex data into auditory experiences1, making it easier for researchers and the public alike to grasp intricate concepts (e.g.,2). Beyond its role in disseminating scientific findings, sonification also serves as a crucial assistive technology for visually impaired individuals, allowing them to analyze data through auditory means (e.g.,3,4). Sonification might even help revealing unrecognized patterns and feedbacks in unwieldy datasets5. A wide number of sonification methods and algorithms have been developed and applied in the context of astronomy6–17 and other fields (e.g.,18,19). Recently, Artificial Intelligence algorithms based on Neural Networks and Deep Learning have also been implemented20,21. Enge et al.22 present an overview of the field of sonification, analyzing the complementarity and redundancy between sonification and visualization. Recent studies even suggest that data sonification has emerged as a viable alternative to data visualization23. It has been shown that audification as a tool for the spectral analysis of time-series data24 can be particularly useful in the presence of low signal-to-noise ratios, where smaller biases are obtained with auditory than with visual stimuli25. For long, it has been suggested to use the loudness, pitch, and duration of each instrumental tone to obtain a multi-dimensional perceptual scaling of musical timbres26. In fact, sonification can also be used to represent additional dimensions in complex multi-dimensional datasets27–29. Audification and sonification of texture in images, have also been proposed30 to represent spatial data, for example, in the context of geographical information31,32, and biomedicine33,34. However, despite its growing popularity and utility as an additional tool to visualisation35, current sonification methods are primarily limited to one-dimensional data due to the lack of practical and robust methods for quantitatively handling multi-dimensional data with solely auditory stimuli. This limitation significantly restricts the range of applications and the depth of insight that sonification can provide. To address this gap, there is a need for innovative approaches that can effectively capture the complexity of multi-dimensional datasets and translate them into meaningful quantitative auditory representations. One discipline poised to greatly benefit from sonification methods capable of handling multi-dimensional data is neuroimaging. The primary goal of this field is to extract meaningful information from complex three-dimensional (3D) and four-dimensional (4D) images that capture both the structure (3D) and dynamics (4D, with time as the fourth dimension) of the human brain, recorded through tomographic techniques such as X-ray computed tomography (CT), positron emission tomography (PET), or magnetic resonance imaging (MRI). In MRI, for instance, highly detailed 3D images of brain structures are obtained with excellent spatial resolution–typically using cubic voxels measuring mm–where voxel intensity reflects the type and density of tissue. While these images are commonly used in clinical settings to detect abnormalities or injuries, advanced mathematical techniques in neuroimaging research allow for the quantification of these images and the identification of subtle differences between groups (e.g., healthy versus pathological subjects or across different age groups) or within the same individual over time. These differences are often not visually apparent due to the complexity of the 3D data, necessitating the use of sophisticated analytical tools to uncover them (see, e.g.,36). Sonification could offer an innovative approach to making these hidden patterns more accessible and interpretable. Multi-dimensional sonification In this study, as part of the CosmicBrain project1, we propose a novel sonification method that leverages higher-order statistics in Fourier space to construct one-dimensional arrays from multi-dimensional data and apply it to human brain magnetic resonance imaging (MRI). A sketch of the methodology introduced in this work can be found in Fig. 1). The details of the different steps are described below and in the method’s section. This diagram illustrates the sonification method for three-dimensional data, utilizing higher-order statistical analysis in Fourier space. The two-point statistics are represented by sticks of varying lengths, which connect pairs of voxels (in the configuration space analog), indicating their respective intensities from the MRI scan. The three-point statistics are depicted through different triangle configurations, where two fixed side lengths are considered with varying subtended angles to connect three voxels (in the configuration space analog). A one-dimensional function is derived by combining these statistical measures, which can then be directly sonified. The range of MIDI notes was adjusted to accommodate the signal with the widest dynamic range (the younger population represented by green circles). Consequently, the older populations (represented by white circles) were mapped using this reference, ensuring a fair basis for comparison. Multi-dimensional clustering analysis based on higher-order statistics The analysis of three-dimensional spatial distributions is a well-established approach in cosmological large-scale structure studies, particularly for understanding the clustering properties of complex systems. In this context, we focus on using two-point and three-point correlation functions to probe the underlying structure of the data. The two-point function captures pairwise correlations, but the three-point function extends this to higher-order statistics. It was first computed in configuration space in the 1970s from the three-dimensional galaxy distribution37, and later extended to Fourier space as the bispectrum38. The three-point statistics are precious in cosmological studies for understanding structure formation39 and for investigating the bias between galaxies and dark matter tracers40 as it is sensitive to non-Gaussianities induced by gravitational evolution. These statistics are sensitive to the complex, multi-dimensional patterns in the data, capturing information beyond what two-point statistics can reveal41. By combining two- and three-point statistics, we gain a more comprehensive understanding of the clustering properties of structures, an approach that has been successfully applied to galaxy survey data (see42). In particular, the three-point function measures non-Gaussian features that are key for understanding the cosmic web’s morphology39. Higher-order statistical analysis applied to magnetic resonance imaging data Using three-dimensional MRI data, our study leverages the previously mentioned well-established statistical tools to characterize the clustering of brain structures. This work is part of the CosmicBrain project, where techniques inspired by cosmological analysis have been adapted to investigate the human brain, with the broader goal of assessing whether such methods might aid in the early detection of neurodegenerative diseases. The cosmic density field represents a multivariate statistical problem, where each voxel in the three-dimensional grid provides statistical information about the cosmic volume44. Similarly, in the context of brain MRI, each voxel represents a location within the brain’s anatomy, allowing us to capture intricate spatial distributions. By applying two- and three-point statistics to brain MRI data in Fourier space, we transform this complex three-dimensional information into a more interpretable format, such as the clustering amplitude of a triangle configuration given two sides as a function of the subtended angle. This kind of data can be trivially converted into time-series data for sonification. The combination of two- and three-point statistical analysis allows for a robust framework that captures the essential structural characteristics of the data, opening up new avenues for scientific inquiry and improving data accessibility. Details on the sonification method can be found in the method’s section. Neuroimaging-specific auditory mapping Based on the OASIS-3 database (see methods section), we examine the bispectrum configuration of the MRI data from healthy subjects in different age ranges, which exhibits complex behavior by integrating information from both the two- and three-point statistics, commonly referred to as the reduced bispectrum. In configuration space, the calculation of the three-point statistics implies the computation of all possible triplets of voxel intensities and evaluating their joint correlations, with standing for the intensity and being the voxel coordinate in the three-dimensional mesh. That computation is combinatorially expensive, scaling with for N voxels. By exploiting the convolution theorem and the Fourier representation, one can instead compute the raw bispectrum as , where is the Fourier transform of the signal with k-vector , and we have used the triangle closure condition: . This avoids explicitly enumerating triplets in configuration space; instead, one needs FFTs and pointwise products in frequency space, reducing the computing cost to about per FFT and linear operations. The reduced bispectrum considered in this study (configuration Q019036) is computed by dividing the raw bispectrum with the hierarchical one: , where P(k) is the power spectrum and the k-mode. For a comprehensive introduction to higher-order statistics and reduced bispectra in particular, and MRI preprocessing and harmonization steps, we refer to Carnero-Rosell et al. 2543 and references therein. Specifically, we focus on the reduced bispectrum configuration Q019036 with wave numbers and mm (see Fig. 2). In the present work, we concentrate on its sonification process. However, to highlight the strength of the chosen configuration, we show the age regression using neural networks based on the power spectrum (two-point statistics in Fourier space), the reduced bispectrum, separately (see Fig. 3). Details on the neural networks architecture can be found in Carnero-Rosell et al. 2543. Sonification of bispectra from human MRI, using an age group of 80-100 years as an example. The original data and the corresponding inverse mapping after sonification are displayed, with the bispectrum for the 40-50 age group included as a reference. The ratios and differences between the original and the inverse-mapped sonified signals are also presented. The x-axis represents the subtended angle between two sides with wave numbers and mm, which is interpreted as the time translating angles to seconds within the sonification process. Regression based on individual summary statistical measures applied on the OASIS-3 dataset using neural networks indicating the MAE and r-Pearson correlation coefficient using (left:) the P(k) of the intensities; (right:) the reduced bispectrum for configuration Q01936. Importantly, exploring additional bispectrum configurations is beneficial, as it allows capturing information across multiple spatial scales (both large and small), thus enabling a more comprehensive characterization of brain aging. We selected the configuration above as a challenging example since it spans a wide range of values yet demonstrates subtle internal variations, which are particularly difficult to represent effectively through sonification. Nonetheless, for the sake of completeness, we provide sonification for the additional three raw bispectrum configurations used in Carnero-Rosell et al. 2543 (B010041,B030031,B080116) to achieve an optimal age regression (see end of the method’s section). Our goal is to assign distinct musical notes and envelope to the significant variations in the bispectrum signal, allowing for a smooth auditory representation of its transitions, while ensuring the signal remains within the audible frequency range for adult human listeners. This task becomes challenging, when realising the large bispectrum range differences between the younger and older groups. To achieve this, we have selected a range of notes spanning eight octaves, including semitones and, when necessary, quarter tones, covering frequencies from a few tens of Hertz up to approximately 8 kHz. Although further tone subdivisions could be explored, they risk impairing auditory distinguishability and should therefore be examined in future work. The frequency range was chosen to encompass both the subtle and significant variations in the bispectrum signal, ensuring that the transitions are captured smoothly and remain within the audible range for adult humans. For more details, see Fig. 4 and methods section, where the specific frequency mapping and note selection are discussed in depth. This setup has allowed us to accurately sonify the bispectra across different age groups while maintaining a consistent amplitude range. Technical details of the sonification procedure can be found in the methods section. We based the amplitude range on the group exhibiting the largest variations, specifically the youngest cohort in our sample: 40-50 years (see left panel in Fig. 2). By standardizing the amplitude in this way, we ensure that the sonification of bispectrum variations remains comparable across age groups, providing a clear auditory representation of the differences in the bispectrum signal across the age spectrum. Sonification piano: notes used in this study covering the frequency range within the sensitivity of adult humans (20-30 yrs: 16, 40 yrs: 14, 50 yrs: 12, 60 yrs: 10, 70 yrs: 8 kHz, see, e.g.,45,46). Regular notes ( for different octaves i ranging from 1 to 8) have been split into two columns (first two on the left) for visualisation purposes. The same has been done with quarter tones (indicated by ) on the right. Semitones are indicated in the third column on the left. The piano convention of black keys for semitones is adopted. We consider sonification cases with and without quarter tones depending on the bispectrum configuration and the desired accuracy. To assess the accuracy of the sonification procedure, we adopt an information-theoretic approach to evaluate how well the original signal can be recovered after undergoing sonification (see methods section). This process involves discretizing the continuous bispectrum signal into integer values, and the challenge lies in determining how much information is lost in this transformation. By analyzing the fidelity of the recovered signal compared to the original, we can quantify the effectiveness of the sonification process and ensure that the discretization does not significantly degrade the essential features of the bispectrum signal. We treat the originally measured bispectrum from MRI data and its sonified version in two ways: first, as a “ground truth” function and a “modeled” function , and second, as their corresponding probability distribution functions and , respectively. The accuracy of the sonification method is then evaluated through metrics defined between and , or between and , allowing us to measure the fidelity of the sonified signal compared to the original. During the reconstruction study, we found that the discretization step can introduce systematic biases, which could affect the accuracy of the sonified signal. However, applying an appropriate rounding procedure, can effectively mitigate these biases. While the differences between distant age groups are perceptible in the sonifications, distinguishing between closer age groups becomes more challenging. For instance, it is difficult to differentiate between the 40-50 and 50-60 year age groups. This is due to the more subtle variations in the bispectrum signal for these adjacent age ranges, which are less pronounced in the auditory representation than the differences observed between more distant groups. To address this challenge, we propose sonifying the differences between different age groups’ bispectrum signals, rather than the signals themselves (see right panel in Fig. 2). This approach emphasizes the subtle variations that might otherwise be hard to distinguish. Additionally, to further enhance the sonified output, we introduce a normalization factor (less than 1) that amplifies these differences, making them more perceptible. This normalization highlights the variations and ensures that the signal remains within a consistent and manageable auditory range. Another way to enhance the sonified representation of the bispectrum signal involves applying a nonlinear transformation that stretches or compresses the dynamic range of the data. This technique improves the dynamic representation of the sonified signal by amplifying or reducing the variations in amplitude, making subtle differences more perceptible, without altering the overall structure or rank order of the data. Applying this technique we have found ways to reduce the ratio deviations (see Fig. 12). Sonification of difference between reduced bispectra (configuration and ) from human brain magnetic resonance imaging for different ages groups with respect to the youngest age group in our sample (40-50 yrs). Upper-left panel: , ; upper-right panel: , ; lower-left panel: , ; lower-right panel: , . The normalization factor shows that the difference signal for the adjacent age groups (50–60 vs. 40–50 yrs) is approximately four times weaker than that for the distant groups (80–100 vs. 40–50 yrs). Because the signal spans both negative and positive values, normalization serves to emphasize contrasts by enhancing the relative prominence of positive versus negative deviations, thereby amplifying the perceptual differences in the sonification. In some cases, we found it necessary to further accentuate these contrasts through the application of a nonlinear transformation. The upper-right and lower-left panels employ the same nonlinear transformation but different normalization factors, resulting in highly comparable signal shapes despite differences in overall scaling. By applying this approach to the differences between the age groups 40–50 and 50–60, 40–50 and 70–80, as well as 40–50 and 80–100, we observe surprisingly similar signal shapes across these comparisons, despite amplitude differences. This suggests the presence of bispectrum configurations that, after appropriate normalization, exhibit shape similarity across age groups, hinting at a form of age-invariant structural patterning. The amplitude variation can be effectively managed with adaptive normalization, ensuring that the differences remain perceptible while maintaining a consistent volume range for the sonified signals. Based on the results shown in Fig. 2 we write the corresponding scores with tones and semitones. The scores corresponding to the sonified bispectrum signals for the 80–100-year and 40–50-year age groups are shown in Fig. 5. The score in Fig. 6 highlights the structured patterns in the differences between these age groups, emphasizing the informational content underlying these variations. It should be noted that statistical tests to assess the perceptual or predictive separability of groups via sonified signals still need to be carried out in future work. Scores corresponding to (top:) the age group of 80 to 100 yrs using as the reference the group of 40 to 50 yrs to set the range; and (bottom:) the age group of 40 to 50 yrs. Scores corresponding to the difference between the age group of 80 to 100 yrs with the group of 40 to 50 yrs. The difference bispectrum signal is obtained by using the strongest signal–namely, the youngest age group in the sample–as the reference, and subtracting from it the bispectrum of the comparison group (here, the oldest age group for this score). Sonification is then applied as for the absolute bispectrum signals. Nonetheless, to illustrate the bispectrum behaviour of individual subjects relative to the group median, we present representative bispectra for all age groups, including an example from a subject with and without diagnosed dementia (Clinical Dementia Rating, cdr > 1) according to the OASIS-3 database (see Fig. 7). Future work might explore sonification with higher order statistics for early detection– identifying bispectral markers that anticipate CDR diagnosis–yet this remains speculative and will require large, longitudinal studies, harmonized preprocessing, and external replication to establish robustness and clinical utility. Bispectra corresponding to configuration Q019036 for different age groups (from left to right: 50-60, 60-70, 70-80, 80-100 yrs). For each group, the subject with the closest bispectrum to the median (medoid: in dashed blue) and one subject with diagnosed dementia (: in red) are displayed. The 40-50 yrs group is omitted, as no subject in this age range was diagnosed with dementia. The green shaded regions represent the contours estimated as the standard deviation of bispectrum values across all subjects within each age bin. These contours indicate across-subject dispersion within each age group, rather than within-subject or temporal variability. For reproducibility, all bispectrum data corresponding to the OASIS-3 database are publicly available (see data availability statement section). Discussion In this work, we investigated the use of sonification of higher-order summary statistics as a method for characterizing complex multi-dimensional patterns within one-dimensional time series. In particular, we explored sonification of bispectrum signals derived from MRI data, with the goal to provide an intuitive and effective auditory representation of complex higher-order statistics, such as the reduced bispectrum. If validated through rigorous clinical studies, such methods could potentially contribute to diagnostic applications, including the detection of early signs of dementia and the monitoring of deviations from expected bispectrum patterns in specific age groups. Nonetheless, these possibilities remain speculative and warrant extensive further research. We have introduced a large notes array, offering a densely distributed set of notes to capture and sonify subtle differences in the bispectrum signal. This array spans a wide frequency range, allowing for the representation of both fine details and broader variations in the data. By ensuring a sufficient number of notes across the audible spectrum, we can maintain precision in the sonification process, making even minor deviations perceptible while preserving the overall integrity of the signal. This approach ensures that the full range of bispectrum variations is accurately conveyed through sound. Additionally, we have demonstrated how to enhance the bispectrum signal for sonification through appropriate normalization, ensuring that amplitude variations remain within a manageable and perceptible range. Furthermore, we have introduced nonlinear rank-ordered transformations to reduce deviations, allowing for dynamic adjustments that stretch or compress the signal’s range while preserving the relative structure and order of the data. These techniques collectively improve the accuracy and clarity of the sonified signal, making subtle differences more audible and enhancing the overall effectiveness of the sonification process. Limitations of higher-order sonification However, one should also consider that this method has clear limitations. First, the perceptual interpretation of sonified signals is highly dependent on the listener’s auditory acuity, experience, and training, introducing variability that does not arise in purely visual analysis. Second, the complexity of mapping multi-dimensional statistical structures onto sound raises the risk of oversimplification or perceptual overload if not carefully designed, especially when considering adjacent age groups. Third, while one bispectrum configuration captures non-linear interactions in the data, it is only one of many possible higher-order statistics; hence, the diagnostic value of bispectrum sonification must be validated against established biomarkers and complementary methods including a robust statistical inference study. Moreover, without systematic user studies–including both clinical experts and lay participants–the extent to which sonification enhances sensitivity or diagnostic accuracy remains unknown. Finally, although auditory methods can be powerful, they are unlikely to replace visual analysis; rather, their strength may lie in serving as an auxiliary tool that reveals patterns as a preliminary step. Conclusions Future work might explore whether sonifying bispectrum difference signals could serve as a useful method in acoustic diagnostics, particularly for identifying deviations from a reference bispectrum signal across different age groups. By focusing on the differences between bispectrum signals, subtle variations –that might hypothetically indicate cognitive changes– can be identified. This approach also helps to distinguish among close age groups, although in those cases the differences show more noisy patterns, as illustrated in Fig. 12. The extension of this study to the sonification of bispectrum signal configurations across multiple scales can significantly enhance the signal-to-noise ratio. While this approach shows promise, a thorough study investigating its full potential for medical diagnostics, particularly in neurodegenerative diseases, is left for future work. Beyond accessibility, the technique may enable visually impaired researchers and clinicians to analyze complex multi-dimensional data by attending to subtle auditory cues. Exploring large sets of bispectrum configurations in this way could reveal patterns otherwise difficult to perceive, and–speculatively–linear or nonlinear combinations of configurations might summarize functional dependencies relevant to brain aging and pathology. For instance, a visually impaired physician might recognize distinctive auditory features that suggest closer inspection of specific MRI scans, and with training, the ability to focus on a single sense and filter out noise could further heighten sensitivity to such patterns. Although visual plots are often apprehended more quickly, sonification offers complementary advantages, allowing physicians to cultivate auditory expertise and even integrate data listening into multitasking contexts, such as commuting or routine walks through a hospital. Since the bispectrum vanishes for Gaussian-distributed signals, departures from Gaussianity can be detected through bispectral analysis. The higher-order sonification technique described here provides a general mathematical framework that could, in principle, be extended to other multi-dimensional, non-Gaussian datasets. Nevertheless, such extensions would require domain-specific methodological adjustments and empirical validation beyond the structural brain MRI data analyzed in this study. Acknowledgements The research conducted in this work was primarily funded by the the Cabildo de Tenerife under IACTEC Technological Training Program, grant TF INNOVA, in support of the Cosmic Brain project starting in 2021 (PI: FSK). EPK and FSK thank Matt Russo for sharing the very instructive sonification instructions in python7, which have been followed as a basis for the code developed in this article. The authors acknowledge the Spanish Ministry of Science and Innovation for financing the Big Data of the Cosmic Web project: PID2020-120612GB-I00/AEI/10.13039/501100011033 (PI: FSK) and the Cosmology with Large Scale Structure Probes project funded by the IAC (PI: FSK), which provided numerical tools and computing facilities. EPK thanks the IAC for hospitality during her internship in the summer 2024. FSK and EPK thank Dr María Joyanes Pérez for discussions on musical notes. The authors thank Jorge Andrés Pérez Prieto and Nicola Caon for their support preparing the Cosmic Brain web-application www.cosmic-brain.org. Methods Input data The data considered in this study was prepared by Carnero-Rosell et al. 2543 within the Cosmic Brain project based on the OASIS-3 database47. All MRI data was collected through the Knight Alzheimer Research Imaging Program at Washington University in St. Louis, MO, USA. Some of the MRI data was collected on a Siemens Vision 1.5T, while the majority of the scans came from two different versions of a Siemens TIM Trio 3T (Siemens Medical Solutions USA, Inc). Participants were lying in the scanner in a supine position, head motion was minimized by inserting foam pads between the participant’s head and antenna coil, and for some participants a vitamin E capsule was placed over the left temple to mark lateralization. A 16 channel head coil was used in all scans. Although a variety of different structural and functional imaging protocols are included in the OASIS-3 dataset such as FLAIR, DTI and ASL, here we focused on the T1w scans. The T1w images were acquired using a 3DMPRAGE protocol TI/TR/TE: 1000/2400/3.08 ms, flip angle = 8, resulting in 1 mm isotropic voxels. Data were provided by OASIS-3: Longitudinal Multimodal Neuroimaging: Principal Investigators: T. Benzinger, D. Marcus, J. Morris; NIH P30 AG066444, P50 AG00561, P30 NS09857781, P01 AG026276, P01 AG003991, R01 AG043434, UL1 TR000448, R01 EB009352. AV-45 doses were provided by Avid Radiopharmaceuticals, a wholly owned subsidiary of Eli Lilly. We verified that only approximately 3% of the OASIS-3 dataset was acquired using 1.5T MRI scanners. Excluding this subset from our analysis results in negligible to no impact on the overall results. Sonification frequency range The first note has a frequency of 32.702 Hz, while the last regular note has a frequency of 7902.13 Hz. In case we include quarter tones, we need to make additional calculations. The 12-tone equal temperament system divides an octave into 12 equal parts, where each part corresponds to a semitone. In this system, the frequency of a note increases by a factor of 2 when going up by an octave. Moving up by one semitone corresponds to multiplying the frequency by , since there are 12 semitones in an octave. The cent is a unit used to measure musical intervals, specifically fractions of semitones. There are 100 cents in a semitone. To calculate the frequency change for a given number of cents, we use the formula , where n is the number of cents. For a quarter tone, which is 25 cents (half of a semitone), we substitute into the formula: . To calculate the frequency of (25 cents above ), we have to multiply the original frequency by the quarter tone factor: Hz. Hence, the total frequency range we are considering when including quarter tones spans from 32.702 to 7961.09 Hz, just below the typical upper sensitivity limit of 8 kHz for humans of about 70 yrs. This should be considered when targeting older age groups. Methodology overview A summary is presented in the flow diagram in Fig. 8. This methodology section describes the steps involved in converting bispectrum MRI data into sonifiable MIDI sequences. The sonification steps can be seen in Fig. 9. Figure 10 shows the sonification procedure to present relative signals with respect to a reference one which determines the range. We can see from the upper-left panel of Fig. 10 (corresponding to the bispectra shown in Fig. 11), that the differences between the 40-50 and 50-60 years group are very tiny. In such cases it seems more adequate to focus on the sonification of the differences. We demonstrate in Fig. 12 how those differences can be enhanced applying the method explained below. MRI data preparation: The first step involves obtaining the MRI brain intensity map on a mesh. This map is represented in three dimensions, with each voxel containing an intensity value corresponding to a specific brain location. These values provide the base data for subsequent analysis. MRI data preparation: The first step involves obtaining the MRI brain intensity map on a mesh. This map is represented in three dimensions, with each voxel containing an intensity value corresponding to a specific brain location. These values provide the base data for subsequent analysis. Statistical analysis: To capture the structural complexity of the brain, we compute two-point and three-point correlation functions in Fourier space. Details on these computations can be found in Carnero-Rosell et al. 2543 and references therein. Statistical analysis: To capture the structural complexity of the brain, we compute two-point and three-point correlation functions in Fourier space. Details on these computations can be found in Carnero-Rosell et al. 2543 and references therein. Bispectrum normalization: The data is normalized to ensure consistency across datasets: , where is a predefined normalization constant. This step ensures that intensity values are comparable and prevents skewing due to extreme values. Bispectrum normalization: The data is normalized to ensure consistency across datasets: , where is a predefined normalization constant. This step ensures that intensity values are comparable and prevents skewing due to extreme values. Bispectrum linear mapping to a positive definite range: We define the linear mapping function “”: The normalized data is linearly mapped to a positive finite range to facilitate MIDI conversion: Bispectrum linear mapping to a positive definite range: We define the linear mapping function “”: The normalized data is linearly mapped to a positive finite range to facilitate MIDI conversion: Time linear mapping to a positive definite range: The subtended angles from triangle configurations are converted to a time line: We are choosing, per default, a duration of 2 seconds, a duration_beats of 20, and a beats per minute (bpm) of 60. Time linear mapping to a positive definite range: The subtended angles from triangle configurations are converted to a time line: We are choosing, per default, a duration of 2 seconds, a duration_beats of 20, and a beats per minute (bpm) of 60. Nonlinear rank-ordered transformation: To enhance perceptibility, the data is then exponentiated by a factor : . This process stretches () or compresses () the dynamic range of the data, improving the sonification’s dynamic representation, while preserving the structure and rank order. For we have . Nonlinear rank-ordered transformation: To enhance perceptibility, the data is then exponentiated by a factor : . This process stretches () or compresses () the dynamic range of the data, improving the sonification’s dynamic representation, while preserving the structure and rank order. For we have . Sonification: mapping to MIDI notes: The transformed intensity values are mapped to MIDI2 notes using the function: where f is the frequency corresponding to the intensity value. The MIDI note number 69 corresponds to the musical note A4, which is standardized at a frequency of 440 Hz (often referred to as “concert pitch”). Quarter tones and octave shifts are adjusted accordingly. The MIDI note pitch is derived by discretizing the intensity range into predefined musical notes (see Fig. 4): where the range is adjusted, in the relative representation case (see Fig. 10), to a reference signal covering the largest range: and The scale factor is set to one when computing differences. With the index we obtain the MIDI notes: . The volume (or velocity) of each note is assigned on a suitable range: where the velocity lies within the range to ensure perceptibility. The term “velocity” in MIDI was chosen to represent the intensity of a note, but it doesn’t refer to the literal speed of a note’s sound. Instead, it refers to how quickly the key is pressed (from a physical perspective) and how that is mapped to the loudness or dynamic quality of the note. Sonification: mapping to MIDI notes: The transformed intensity values are mapped to MIDI2 notes using the function: where f is the frequency corresponding to the intensity value. The MIDI note number 69 corresponds to the musical note A4, which is standardized at a frequency of 440 Hz (often referred to as “concert pitch”). Quarter tones and octave shifts are adjusted accordingly. The MIDI note pitch is derived by discretizing the intensity range into predefined musical notes (see Fig. 4): where the range is adjusted, in the relative representation case (see Fig. 10), to a reference signal covering the largest range: and The scale factor is set to one when computing differences. With the index we obtain the MIDI notes: . The volume (or velocity) of each note is assigned on a suitable range: where the velocity lies within the range to ensure perceptibility. The term “velocity” in MIDI was chosen to represent the intensity of a note, but it doesn’t refer to the literal speed of a note’s sound. Instead, it refers to how quickly the key is pressed (from a physical perspective) and how that is mapped to the loudness or dynamic quality of the note. The reduced bispectra sonification can be seen in Figs. 11 and 12. The latter shows the application of the normalization and the nonlinear transformation. The nonlinear transformation enhances the representation of features that exhibit minimal variations within a signal that spans a wide range of values. We can see in the lower-left panel of Fig. 9, how some bispectrum bins acquire the same value in the sonfication process. This problem is not present in the lower-right panel where was applied. As a consequence of this transformation the ratio plots in the top and bottom panels of Fig. 12 display a flatter behaviour than in Fig. 2. Figures 13 and 14 show the difference bispectra for a consistent transformation across ages with half tones only. Flow diagram of the higher-order sonification pipeline (in green: data processing, as provided by Carnero-Rosell et al. 202543; in blue: higher-order sonification steps introduced in this work; in red: incomplete work which needs further development). After MRI preprocessing and Fourier-space statistics are computed, a bispectrum configuration is selected and a mode is chosen (absolute, or difference relative to a designated reference). The signal is then normalized (), linearly mapped to a finite range, and optionally transformed with a rank-ordered exponent (); in absolute mode, the range is scaled to the reference. Angles are mapped to time, and amplitudes to MIDI pitch/velocity to produce the audio sequence. The lower branch shows the mode choices in more detail and the additional inverse mapping and information-loss evaluation used to assess fidelity. Statistical inference and perceptual testing are not performed here and remain future work. Sonification procedure across four panels: The upper-left panel shows the original signal with each data point represented by a small circle. In this example, the difference of the reduced bispectra (configuration and ) for age groups: 80-100 and 40-50 yrs is shown. In the upper-right panel, the minimum value is subtracted from the signal. The lower-left panel illustrates the signal discretization process, additionally applying a linear mapping to convert the signal into the range of MIDI values that are perceptible to adult human listeners. Finally, the lower-right panel demonstrates the effect of incorporating a nonlinear rank-ordered transformation with a power-law of exponent . The velocity (volume) of the sonified signal is represented through the size of the circles. The normalization is one. By preserving the fundamental patterns yet compressing the dynamic range, the nonlinear transformation facilitates detailed sonification, making subtle variations easier to discern (see lower panels). However, care must be taken in this operation, since the signal is altered and comparisons may become unreliable if different nonlinear transformations are applied across data sets. Consistency in the chosen transformation is therefore essential. Similar sonification procedure to Fig. 9, but showing relative reduced bispectra relations (configuration and ). The different panels show the discretized signal for the 50-60 yrs (upper-left), 60-70 yrs (upper-right), 70-80 yrs (lower-left), 80-100 yrs (lower-right) age group within the range of the reference 40-50 years group mapped to the corresponding MIDI values. The filled circles stand for the reference bispectrum. Sonification of reduced bispectra (configuration and ) from human brain magnetic resonance imaging for different ages. The original data and the corresponding inverse mapping after sonification are indicated for each age range. Same as Fig. 12, but for configurations: left column: (B010041) with (,) and right column: (B030031) with (,). Same as Fig. 12, but configurations: left column: (Q019036) with (,) and right column: (B080116) with (,). Information loss To assess the accuracy of the sonification, the information loss is calculated by transforming the MIDI values back into the original bispectrum space: . This includes reversing the exponentiation (for ): . Information loss is measured using, e.g., the normalized mean squared error (NMSE): For a more complete metric analysis see Table 1 applied to the reduced bispectrum for both absolute and differences sonification. These metrics provide insights into the accuracy of the sonification and guide any necessary adjustments. Table 1Metrics and their mathematical formulae computed for different age groups corresponding to the reduced bispectra (configuration and ). The ideal case is indicated within brackets, e.g., or . “ABS” stands for sonification of absolute bispectra, and “DIFF” stands for difference between bispectra of a group and a reference one of .Statistical Measure Name (SMN)Mathematical FormulaAge GroupMean40-5050-6060-7070-8080-100VarianceABSABSDIFFABSDIFFABSDIFFABSDIFFMean Absolute Error 0.00260.01620.00150.00030.00460.01040.00640.00780.0064Mean Squared Error 0.00040.00040.00000.01690.00000.00010.00010.00010.0001Root Mean Squared Error 0.01940.01920.00200.01690.00620.01270.00890.01020.0088Normalized MSE [%]0.02550.02740.05710.02750.08320.02150.05870.01990.0292Relative MSE [%]0.41380.35100.01320.27630.03510.17340.03030.14220.0217Signal-to-Noise Ratio 3924.63648.11750.33642.71202.44649.41704.25017.83420.7R-squared 0.99970.99970.99940.99970.99930.99980.99940.99980.9997Explained Variance 0.99980.99980.99940.99970.99920.99980.99920.99980.9997Cosine Similarity 1.00001.00000.99991.00000.99991.00000.99991.00000.9999Kullback-Leibler Divergence 0.00000.00000.00010.00000.00010.00000.00010.00000.0001Cross-Entropy 0.05350.05340.05330.05340.05210.05340.05160.05340.0504Total Variation Distance 0.00240.00240.00430.00260.00660.00220.00620.00210.0053Hellinger Distance 0.00270.00260.00420.00290.00600.00250.00550.00220.0052 Metrics and their mathematical formulae computed for different age groups corresponding to the reduced bispectra (configuration and ). The ideal case is indicated within brackets, e.g., or . “ABS” stands for sonification of absolute bispectra, and “DIFF” stands for difference between bispectra of a group and a reference one of . Depending on the metric, we will treat the data as a function (Y) or as a PDF (P): The metrics in Table 1 were obtained with the inclusion of quarter tones. Results without quarter tones were worse by about an order of magnitude in the NMSE. These metrics are intended to evaluate the reconstruction fidelity of the sonification pipeline, rather than to quantify inter-subject variability. The metric values therefore reflect deterministic reconstruction accuracy and not population-level dispersion. All reported metrics are descriptive in nature and no statistical inference or testing of group differences has been performed. Author contributions FSK proposed this project, directed the study, prepared the software and wrote the paper. EPK contributed to the sonification piano, coding, sonification and metric calculations. ACR, NJ, AM and EP contributed to the development of the higher-order statistics analysis method, to providing the input data of this work, and to improving the manuscript. Funding Spanish Ministry of Science and Innovation: PID2020-120612GB-I00/AEI/10.13039/501100011033. IACTEC Technological Training Program, grant TF INNOVA, in support of the Cosmic Brain project. Data availability All the bispectrum data and software used in this work are publicly available at https://github.com/pacoshu/sonification. The software developed for sonification relies on several specialized libraries: Streamlit for the web application interface; music libraries including midiutil and music21 for MIDI manipulation and score generation; statistical libraries such as sklearn.metrics, scipy.stats, and scipy.spatial.distance for analysis and metric calculations; and data processing and visualization libraries, including pandas, numpy, matplotlib, as well as io, os, and itertools for efficient data handling and plotting. A GitHub repository contains all bispectrum calculations used in the brain age regression, together with the corresponding neural network code: https://github.com/aureliocarnero/brainage-higherorder/. Declarations Competing interests The authors declare no competing interests. Footnotes www.cosmic-brain.org MIDI (Musical Instrument Digital Interface) is a technical standard that describes a communication protocol, digital interface, and electrical connectors that connect a wide variety of electronic musical instruments, computers, and related audio devices for playing, editing, and recording music48,49. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement All the bispectrum data and software used in this work are publicly available at https://github.com/pacoshu/sonification. The software developed for sonification relies on several specialized libraries: Streamlit for the web application interface; music libraries including midiutil and music21 for MIDI manipulation and score generation; statistical libraries such as sklearn.metrics, scipy.stats, and scipy.spatial.distance for analysis and metric calculations; and data processing and visualization libraries, including pandas, numpy, matplotlib, as well as io, os, and itertools for efficient data handling and plotting. A GitHub repository contains all bispectrum calculations used in the brain age regression, together with the corresponding neural network code: https://github.com/aureliocarnero/brainage-higherorder/.