Proton Nuclear Magnetic Resonance With Time‐Frequency Analysis: A Potential Diagnostic Approach for Keloids

ABSTRACT

Keloids are chronic fibroproliferative skin disorders with high recurrence rates and limited treatment options, yet reliable diagnostic biomarkers are lacking. Current classification systems rely heavily on clinical observation, underscoring the need for objective, noninvasive tools. In this exploratory study, serum‐based ¹H nuclear magnetic resonance (NMR) measurement combined with short‐time Fourier transform (STFT) for time‐frequency analysis was performed, followed by principal component analysis (PCA), to investigate potential patient subgroups. Serum samples from 29 patients were analysed and PC1 scores suggested two potential patient subgroups. Retrospective analysis showed that these subgroups differed primarily in keloid aetiology: one group predominantly included cases arising from unclear or minimal causes (e.g., acne, folliculitis), whereas the other comprised cases following clear traumatic events (e.g., surgery). Although most clinical variables showed no significant differences, significant differences in aetiology and Japan Scar Workshop Scar Scale (JSS) scores support the biological relevance of this separation of subgroups. These findings suggest that the time‐frequency features of NMR signals from serum samples capture systemic characteristics associated with keloid pathophysiology. If validated in larger cohorts, this approach may serve as a noninvasive adjunct to clinical assessment and lay the foundation for objective patient stratification and precision‐guided treatment strategies.

Key Points

• Keloid diagnosis and classification rely largely on clinical observation.

• Time‐frequency analysis of serum ¹H‐NMR data revealed latent subgroups among keloid patients.

• Subgroup separation was most strongly associated with keloid aetiology.

• This approach may provide a noninvasive adjunct for aetiology‐based stratification and personalized treatment.

Abbreviations

BMI

body mass index

CRP

C‐reactive protein

DANTE

Delays Alternating with Nutation for Tailored Exitation

FFT

fast Fourier transform

FID

free induction decay

HS

hypertrophic scar

JSS

Japan Scar Workshop Scar Scale

JSW

Japan Scar Workshop

NIPALS

nonlinear iterative partial least squares

NMR

nuclear magnetic resonance

PCA

principal component analysis

STFT

short‐time Fourier transformation

1. Introduction

Keloids are fibroproliferative skin disorders that arise from abnormal wound healing and are characterised by excessive extracellular matrix deposition that extends beyond the original injury site. They often cause significant aesthetic and functional impairment, including pruritus, pain and restricted movement [1, 2].

Keloids can develop from minor or unrecognised skin injuries (e.g., acne, folliculitis, insect bites), but they may also occur following clearly identifiable traumatic events such as surgical incisions. A hallmark feature of keloids is uncontrolled fibrous proliferation beyond wound boundaries [3]. In contrast, hypertrophic scars (HSs) typically develop following well‐defined injuries and remain confined to the original wound site, often improving over time. Unlike HSs, keloids are notably refractory to treatment and prone to recurrence, often requiring multimodal management including surgery, steroid therapy and radiotherapy.

Currently, the diagnosis and severity assessment of keloids rely heavily on clinical observation due to the absence of specific laboratory tests [4]. Although histological features such as keloidal collagen can aid diagnosis, biopsy is invasive and sometimes inconsistent with clinical findings [5, 6]. To address these limitations, the Japan Scar Workshop Scar Scale (JSS), based on 12 clinical and genetic risk factors, is widely used for scar assessment [7, 8]. It classifies scars into mature scars (0–5 points), hypertrophic scars (6–15 points) and keloids (16–25 points). Keloid cases can be further subdivided into moderate (16–19 points) and severe (20–25 points) forms, depending on treatment responsiveness. Though useful for classification, the JSS depends on the presence of existing lesions and may not fully reflect systemic predisposition (detailed criteria are provided in Table S1).

Clinically, keloids that arise following minor or clinically indistinct inflammatory stimuli, such as acne or folliculitis, may reflect a stronger intrinsic predisposition and often require more intensive treatment strategies and closer follow‐up. In contrast, keloids that develop after clearly identifiable trauma, such as surgery, may be managed more effectively through cause‐oriented preventive measures and postoperative care. Therefore, identifying aetiology‐based subtypes of keloids may help predict treatment resistance and recurrence risk and support more appropriate selection of therapeutic strategies.

Increasing evidence supports the idea that keloid pathogenesis involves systemic dysregulation, including genetic susceptibility, abnormal immune responses and altered metabolism [9, 10, 11, 12, 13]. Serum‐based omics approaches, including proteomics, transcriptomics and metabolomics, have identified potential molecular markers such as NEDD4 variants, circular RNAs and microRNAs, but their diagnostic utility remains limited [14, 15, 16, 17, 18, 19].

Frequency analysis of proton nuclear magnetic resonance (¹H NMR) signals, namely, spectroscopy, shows detailed spectral features of biological samples [20, 21]. When combined with time‐frequency transformation of NMR signals (free induction decay (FID)), such as via short‐time Fourier transform (STFT), this technique captures decay and frequency characteristics. Previous studies have shown that time‐frequency analysis of NMR‐FID signals improves diagnostic resolution compared with conventional frequency analysis [22, 23, 24, 25, 26, 27, 28, 29].

Despite these advances, the application of serum‐based NMR and time‐frequency analysis to keloids remains unexplored. In particular, whether distinct subgroups of keloids can be objectively identified based on serum characteristics is unknown.

In this study, we hypothesized that STFT‐based analysis of ¹H NMR serum signals, combined with principal component analysis (PCA), can suggest naturally occurring subgroups among keloid patients and investigated their associations with clinical characteristics. The ultimate goal was to explore whether this approach could serve as a foundation for noninvasive, objective stratification of keloid patients to guide personalised treatment strategies.

2. Methods

2.1. Study Design and Ethics

This prospective, case–control study was conducted at Nippon Medical School Hospital, Tokyo, Japan, in accordance with the Declaration of Helsinki. The study was approved by the ethics committee at Nippon Medical School (Approval No: B‐2020‐194). Written, informed consent was obtained from all participants.

2.2. Patient Recruitment

Between December 2019 and October 2022, patients scheduled for surgical excision of pathological scars at Nippon Medical School Hospital were screened for eligibility. Patients clinically diagnosed with keloids (i.e., raised scars extending beyond original wound borders) were recruited. Cases with keloids limited to the ear were excluded because these lesions often result from repeated piercing trauma and may represent a distinct anatomical and pathophysiological entity. Patients with ear keloids who also had keloids in other locations were included. Ultimately, 29 patients were enrolled for serum sampling and subsequent ¹H‐NMR analysis (Figure 1).

FIGURE 1.

Flowchart of patient recruitment and inclusion for serum NMR analysis. Flowchart of patient recruitment and inclusion for serum ¹H‐NMR analysis. Of 54 patients initially screened, 25 with keloids limited to the ear were excluded, leaving 29 patients for NMR‐based analysis.

2.3. Data Collected and Classification of Case Patients

Demographic and clinical data were obtained from prospectively maintained medical records. Collected variables included age, sex, body mass index (BMI), aetiology of keloid (unclear/minor cause vs. clear trauma), family history, C‐reactive protein (CRP) positivity and the JSS score [7, 8] (see Table S1 for scoring details).

2.4. Serum Samples

Peripheral blood samples were collected on admission for scar treatment. Samples were centrifuged at 3000 × g for 5 min and the supernatant was aliquoted and stored at −80°C until analysis.

2.5. Acquisition of FIDs From Serum Samples

¹H‐NMR measurement of the serum samples and STFT‐based FID analysis were conducted as previously described [30]. Briefly, each serum sample (100 μL) was mixed with 500 μL deuterium oxide (²H₂O) (ISOTEC, Sigma‐Aldrich, St. Louis, MO, USA) and pipetted into a 5‐mm (outer diameter) NMR tube (Wilmad‐Lab Glass, Vineland, NJ, USA) for ¹H‐NMR measurement. NMR measurements were performed at a proton‐resonance frequency of 400 MHz (9.4 T) using an ECZ NMR spectrometer (JEOL Ltd., Tokyo, Japan) interfaced with a probe (digital auto‐tunable type) and equipped with Delta NMR processing and control software (version 5.3.2; JEOL Ltd.). The field was locked to the ²H resonance of the ²H₂O solvent. One‐dimensional ¹H NMR signals were automatically acquired at a probe temperature of 30°C using the programme supplied by JEOL that supported the macro function in Delta. FIDs were acquired using a single pulse with a 2.0 s relaxation delay between repeated pulse sequences. The strong signal arising from free water was suppressed using a DANTE (Delays Alternating with Nutation for Tailored Exitation) presaturation pulse train. Other conditions were as follows: flip angle of NMR pulses, 45 degrees; acquisition time, 1.636 s; repetition time, 3.636 s; spectral width, 10 016 Hz; number of data points, 16 384; and number of steady‐state transients, 400. The FID data were saved in JEOL Delta format (JDF) (Figure 2).

FIGURE 2.

Workflow of serum ¹H‐NMR analysis, including FID signal acquisition, STFT and PCA analysis. Workflow of serum ¹H‐NMR analysis, including acquisition of free induction decay (FID) signals, transformation into spectrograms using the short‐time Fourier transform (STFT) and subsequent principal component analysis (PCA). This workflow was adapted from [31], with modifications for the current study.

2.6. NMR Data Analysis

To analyse how the frequency components of FID signals change over time, the STFT was applied as a time‐frequency analysis method [31] (Figure 2). Whereas the standard Fourier transform (FT) provides the frequency content of the entire signal, STFT involves dividing the signal into overlapping segments with short time intervals and applying the Fourier transform to each segment to obtain both time and frequency information. STFT was performed using a self‐made application programme, the ‘STFT tool’, designed for MATLAB (R2023b, The MathWorks Inc., Natick, MA, USA). FID data of each serum sample were imported into the STFT tool, where STFT processing and spectrogram calculations were conducted. The resulting spectrogram data were output as image and ASCII format data. STFT and spectrogram calculations were performed using a previously reported algorithm [32]. The parameters used in this study were as follows: window function, Hanning; FFT block size, 1024; separate points, 512; shift length, 32; and window length, 256. The spectrogram was generated as a visual representation of the time‐frequency analysis. Spectrogram data, output in ASCII format and stored in a two‐dimensional matrix, were read from left to right and row by row, then populated in a single row. In this study, each spectrogram was generated as a 512 × 1024 two‐dimensional matrix, but frequency domains with no detectable signals were excluded to reduce the computation load for calculation. The resulting spectrogram was represented as a 512 × 512 two‐dimensional matrix and then reshaped into a single row of 1 × 262 144. After combining all rows into a separate matrix, (number of samples) × 262 144, the resulting dataset was used for multivariate analysis.

2.7. Multivariate Analysis

For multivariate analysis, PCA was performed using the VEKTOR DIREKTOR software (version 1.0., KAX Group, Sydney, Australia). PCA is a dimensionality reduction method that preserves the maximum variance possible using only explanatory variables. It is widely used in exploratory data analysis to detect outliers and potential subgroups [33]. To achieve this goal, PCA calculates new variables called principal components. These variables are obtained as linear combinations of the original variables. The first principal component (PC1) is calculated to have the maximum possible variance. Next, the second principal component (PC2) is calculated to have the maximum possible variance under the constraint that it is orthogonal to the first principal component. Subsequent principal components are calculated in the same manner. The values of these new variables for the observations are called scores and can be interpreted as the projection of the observations onto the principal components [34]. In this study, PCA was performed to explore the relationships between the principal component scores of each serum sample.

The spectrogram dataset was imported into VEKTOR for PCA. The data were then preprocessed using the standard normal variates (SNV) transformation to standardise variable scales for comparison of each sample. PCA was performed with the nonlinear iterative partial least squares (NIPALS) algorithm and leave‐one‐out cross‐validation. A careful examination of the obtained principal component scores is appropriate for the interpretation of similarities between samples. For this examination, line graphs and histograms can be used, but score plots—scatter plots with two specific principal components as axes—are also widely employed. Furthermore, the analysis of the same plot, colour‐coded with additional available data, can provide further insights [35]. The resulting score plots were drawn using the tools in Excel (Microsoft, Redmond, WA).

For further interpretation, to identify the time‐frequency domain of NMR signals that contribute to positive and negative PC scores, the correlation loading values, which are defined by what linear combination of variables a particular component represents [35], of each PC were plotted on a diagram with the same time‐frequency axes as the spectrogram, referred to as the ‘correlation loading plot’. The correlation loading value data obtained from PCA were exported from VEKTOR in CSV format and displayed as correlation loading plots using a programme written for MATLAB [31].

2.8. Statistical Analysis

To examine the relationships between sample characteristics suggested by PC scores and clinical findings, continuous variables (age, BMI, JSS score) were analysed using the Mann–Whitney U test and categorical variables (sex, aetiology, family history, CRP positivity) were compared using Fisher's exact test. All statistical analyses were conducted using Python 3.0. For comparisons involving more than two groups, the Kruskal–Wallis test was applied for continuous variables and the chi‐squared test was used for categorical variables. A two‐sided p‐value < 0.05 was considered significant.

3. Results

3.1. Patient Cohort

A total of 54 patients diagnosed with keloids and scheduled for surgical excision were screened between December 2019 and October 2022. After excluding cases with keloids limited to the ear, 29 patients were included for serum ¹H‐NMR analysis (Figure 1). The median age of these patients was 44 (range, 20–81) years, with 17 males (59%) and 12 females (41%).

3.2. Spectrogram Generation

Serum samples underwent ¹H‐NMR measurements and the resulting free induction decay (FID) signals were transformed using STFT to generate spectrograms representing the time‐frequency distribution of signal intensity (Figure 3).

FIGURE 3.

Spectrograms of STFT‐processed FID signals from serum samples of keloid patients (n = 29). Spectrograms derived from STFT of FID signals obtained from serum samples of 29 keloid patients. Each panel represents an individual sample, displaying time (x‐axis) against frequency (y‐axis), with colour indicating signal intensity.

3.3. PCA

PCA was applied to the spectrogram dataset of serum samples from keloid patients. Among the two‐dimensional score plots composed of various combinations of PCA scores, the PC1 score plots (PC1/PC2, PC1/PC3, PC1/PC4) showed that the sample groups were gently clustered according to the PC1 score values and could be divided into two potential subgroups (Figure 4A–C). Correlation loading plots of PC1 to PC4 projected onto the spectrogram space (Figure 4D–G) identified key time‐frequency domains contributing to this separation.

FIGURE 4.

Principal component analysis (PCA) and correlation loading plots of spectrogram data (n = 29). The PCA PC1/PC2 score plot is shown in A, PC1/PC3 in B and PC1/PC4 in C. Total variances of PC1, 2, 3 and PC4 are 10.08%, 4.63%, 4.26% and 3.96% of total variance, respectively. From the observation of each score plot, the samples tend to be separated into two groups based on the positive or negative values of the PC1 scores. The score plots are labelled based on the tendency for sample separation. The labelling shows that samples in the G1 subgroup have positive PC1 scores, whereas those in the G2 subgroup have negative PC1 scores: Group 1 (red) and Group 2 (green). Specially, of various principal component combinations evaluated, the PC1/PC4 score plot shows the most sample separation. The principal component 1 correlation loading plot is in D, PC2 in E, PC3 in F and PC4 in G. The red and blue areas indicate the time‐frequency domains that are significantly related to the positive and negative contributions of each principal component. Values closer to +1 are shown in darker red and values closer to −1 are shown in darker blue.

3.4. Clinical Correlation of Potential Subgroups

Since the existence of two groups with different time‐frequency characteristics of NMR signals in patient serum was shown in the PC1 score, the two patient subgroups were designated Group 1 (G1, n = 12) and Group 2 (G2, n = 17). The comparison of the clinical characteristics between these groups is summarised in Table 1; the clinical characteristics were compared between the groups to assess biological relevance. The most prominent difference was observed in keloid aetiology: 91.7% (11/12) of patients in G1 had keloids arising from unclear or minimal causes such as acne or folliculitis, whereas only 41.2% (7/17) in G2 had similar etiologies (p = 0.008). Conversely, keloids resulting from clear traumatic events, such as surgery, were more frequent in G2.

TABLE 1.

Comparison of clinical characteristics between the PCA‐based subgroups.

	Group 1″	Group 2″	p‐Value
	(n = 12)	(n = 17)
Age (years)	41.0 [20.0–81.0]	48.0 [20.0–76.0]	0.610
Sex (male)	7 (58.3%)	10 (58.8%)	1.000
Japanese Scar Scale (JSS)	21.5 [15.0–24.0]	18.0 [5.0–23.0]	0.011
BMI (kg/m2)	26.0 [17.6–29.4]	24.1 [20.5–34.8]	0.773
Causes (unknown or minute)	11 (91.7%)	7 (41.2%)	0.008
Family history	4 (33.3%)	4 (23.5%)	0.683
Elevated CRP	4 (33.3%)	6 (35.3%)	1.000

Note: Comparison of clinical characteristics between PCA‐based clusters (Group 1″, n = 12; Group 2″, n = 17). Variables compared include age, sex, Japan Scar Workshop Scar Scale (JSS), BMI, keloid aetiology (unknown or minimal vs. trauma), family history, and CRP positivity. Fisher's exact test was used for categorical variables and the Mann–Whitney U test was used for continuous variables; p‐values < 0.05 were considered significant.

JSS scores were slightly but significantly higher in Group 1 (p = 0.011), whereas no significant differences were observed for age, sex, BMI, family history and CRP positivity.

When the PCA score plots were colour‐coded according to age and the Japan Scar Workshop Scar Scale (JSS), no clear separation along PC1 was observed, in contrast to keloid aetiology (see Figure S1).

4. Discussion

In this exploratory study, PC1 scores obtained from serum ¹H‐NMR spectrogram data divided keloid patients into two potential subgroups and these potential subgroups were found to be primarily associated with the aetiology of the lesions. Specifically, two clusters differed in whether their keloids originated from unclear or minimal insults (e.g., acne, folliculitis) or from clearly identifiable trauma (e.g., surgery). This suggests that intrinsic differences in serum time‐frequency features of FID, reflecting the overall physicochemical features of serum, may capture underlying pathophysiological heterogeneity in keloids.

A major strength of this approach is its feature extraction, which does not rely on prior classification or subjective labelling. Instead, PCA objectively identified latent structures within spectrogram data and retrospective clinical review demonstrated a relationship with causal background. Therefore, these results could be interpreted as visualisation of systemic differences in serum characteristics between keloid subtypes. The clustering pattern was primarily driven by PC1, suggesting that this principal component captures the most relevant systematic variation associated with keloid aetiology. Though not all clinical variables differed significantly, the association between PCA‐derived subgroups and aetiology was striking, supporting PCA as an exploratory tool for keloid stratification.

Furthermore, it is interesting that the clustering pattern was indicated by PC1. This is because, as mentioned earlier, among the multiple principal components obtained, PC1 retains the largest variability [36]. In this study, this can be interpreted to suggest that the aforementioned variability based on serum time‐frequency features of FID is the largest among the many types of variability present in serum samples.

From a clinical perspective, distinguishing keloids triggered by minimal skin insults from those caused by major trauma is critical, since these subtypes likely involve different mechanisms and therapeutic responsiveness. Integration of serum‐based profiling could enable identification of patients at higher risk of recurrence or treatment resistance and inform personalised management strategies, such as adjusting radiation therapy doses or intensifying adjunctive pharmacotherapy for those with a strong intrinsic predisposition [37, 38]. Such an approach aligns with precision medicine principles and may ultimately improve outcomes and reduce recurrence rates.

Although differences in JSS scores between the two clusters were modest, the trend toward higher scores in the unclear/minor aetiology group suggests partial concordance with clinical severity indices. This supports the concept that time‐frequency profiling of FID may capture systemic susceptibility beyond local lesion characteristics. Consistent with this interpretation, colour‐coding of PCA score plots by age or JSS did not reveal clear separation along PC1, further supporting keloid aetiology as the primary factor associated with the observed serum FID feature. At the same time, it should be emphasised that PCA is essentially an adaptive exploratory method [36]. Rather than inferring causal relationships, PCA is designed to extract the most informative structure from complex datasets by reducing dimensionality, simplifying data description and elucidating relationships among observations and variables [34]. Accordingly, the present PCA‐based findings should be interpreted as exploratory insights into data structure rather than definitive causal explanations.

This study has limitations. First, only serum samples from patients clinically diagnosed with keloids were analysed and no non‐keloid or healthy control samples were included. Therefore, this study was not designed to evaluate the discriminatory ability of the proposed method between keloid and non‐keloid conditions and diagnostic specificity could not be assessed. In addition, formal repeatability or reproducibility testing using repeated NMR measurements was not performed, as this exploratory study was based on previously collected clinical serum samples and the available sample volume and study design did not allow systematic repeated measurements. Additional limitations include the single‐center design, modest sample size and exclusion of patients with isolated ear keloids, which may limit generalizability. Moreover, PCA is inherently exploratory in nature and does not allow causal inference. Future studies with larger, multi‐center cohorts and longitudinal follow‐up are needed to validate these findings and determine whether PCA‐based subtypes exhibit differential treatment responses. Since the potential subgroup identification was based on the PC1 score on the score plots, future studies should use objective clustering algorithms or validation in independent cohorts to confirm these findings. Despite these limitations, it is important to consider the practical feasibility and potential clinical implementation of the proposed approach.

From a practical standpoint, the proposed approach relies on proton NMR measurements, which is already widely available in research hospitals and many analytical laboratories. Importantly, this method is not restricted to high‐field research‐grade systems. Widely available NMR apparatus (e.g., 400 MHz systems), which are commonly installed even in mid‐sized institutions and potentially lower‐field systems with appropriate optimisation, may be sufficient for STFT‐based FID analysis.

The analysis is noninvasive, requires only a small volume of serum and can be performed using residual blood samples obtained during routine clinical testing. In addition, the measurement involves minimal consumables, supporting its feasibility as a centralised or referral‐based laboratory test. With further standardisation, automation and validation, this approach could complement conventional clinical assessment rather than serve as a point‐of‐care tool. If validated, STFT‐based FID analysis may serve as a noninvasive adjunct to conventional evaluation, offering a potential strategy for aetiology‐guided therapy and recurrence risk assessment.

5. Conclusions

This study demonstrates that time‐frequency analysis of serum ¹H‐NMR data can suggest clinically relevant potential subgroups of keloid patients, particularly reflecting lesion aetiology. These findings highlight the potential of serum NMR profiling as a noninvasive, objective tool for patient stratification and as a foundation for personalised management strategies. Future studies with larger cohorts and longitudinal follow‐up are warranted to validate these findings and explore their clinical utility in guiding precision therapy.

Funding

The authors have nothing to report.

Ethics Statement

The study was approved by the ethics committee at Nippon Medical School (Approval No: B‐2020‐194).

Consent

Written informed consent was obtained from all participants.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: PCA score plots colour‐coded by age, Japan Scar Workshop Scar Scale, and keloid aetiology.

Table S1: Overview of the Japan Scar Workshop Scar Scale (JSS).

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.