A new approach for sex prediction by evaluating mandibular arch and canine dimensions with machine-learning classifiers and intraoral scanners (a retrospective study)

Abstract

In circumstances where antemortem information concerning the deceased individual is unavailable, forensic experts prepare biological profiling for unidentified human remains that aids in narrowing the search for identity. Biological profiling includes basic demographic information such as sex, age, stature, and ethnicity. Sex identification is the first and key step in the biological profiling of unidentified human remains, as it effectively reduces potential matches by excluding nearly one-half of the suspected cases and facilitates the subsequent stages. This study was conducted to assess the accuracy of artificial intelligence (AI) in predicting sex by analysing mandibular canine dimensions, mandibular intercanine distance (MICD), and mandibular canine index (MCI) obtained from three-dimensional (3D) digital impressions captured by using an intraoral scanner (IOS). The results of the receiver operating characteristic (ROC) test indicated that mean mandibular canine width (MeanMCW) had the highest sexual dimorphism with the area under the curve (AUC) of 0.912, and the Gaussian Naive Bayes (GNB) classifier demonstrated the highest testing accuracy among all machine learning (ML) models, achieving an accuracy of 92.5%. While the outcomes of this study are promising, further studies are imperative to validate these findings with larger sample sizes in different ethnic populations.

Forensic dentistry has emerged as an essential component within the field of forensic science, assuming a significant role in the identification of deceased individuals who cannot be visually recognized or identified through alternative methods. The process involves a thorough examination and assessment of dental records, which are subsequently utilized in legal proceedings to serve the interests of justice^1–3. Identification in such medicolegal contexts poses considerable challenges, particularly when antemortem records are not accessible and individuals provide misleading or inaccurate information, complicating the attainment of accurate conclusions⁴.

Sex determination plays an essential role in the process of identification especially when antemortem records are not available, as it aids in establishing the biological profile of the deceased individual, thereby facilitating identification⁵. Besides skeletal elements, teeth are regarded as crucial components in sex determination, likely due to the high mineral content that makes them the most durable structure of the human body that can withstand extreme temperature, chemical and mechanical degradation during the postmortem phase. Various methodologies have been documented for sex determination in forensic dentistry, incisors, molars, canines, and mandibular parameters have been extensively employed in studies exploring sexual dimorphism^6–10.

However, Canines are frequently utilized for sex identification, due to their robustness in the oral cavity and morphological sexual dimorphism. Since first proposed by Rao et al.¹¹ the mandibular canine parameters such as MCI, mesiodistal (MD) widths, canine height, and MICD are the most frequently utilized measures in sex determination by researchers, and the results of conventional statistical analysis indicated that mandibular canine parameters have significant sexual dimorphism and can be used for sex estimation, while other studies refuted the use of mandibular canine parameters in sex estimation and suggested that this methodology have low sensitivity and specificity¹².

In the digital era, the convergence of technological advancements with the introduction of IOS and 3D radiography made the dental profession undergo a notable transformation. This state-of-the-art technology has not only brought about a significant change in dentistry but has also expanded its utility to the domain of forensic odontology, particularly for purposes of human identification which assume a pivotal role through sex determination, age estimation, and the comparison of antemortem and postmortem data¹³. Due to advancements in software technologies and the availability of efficient hardware, the utilization of digital forensic investigations is progressively gaining popularity, particularly in scenarios involving mass disasters where a high number of unidentified human remains are found¹⁴.

In clinical dentistry, recent advancements in technology have facilitated the precise reconstruction of an individual’s dental characteristics. For instance, IOS has the capability to generate 3D representations of the oral cavity, allowing for a thorough analysis of treatment efficacy¹⁵. These 3D scans of teeth and soft tissues serve as viable alternatives to conventional plaster models in dental practice¹⁶. Moreover, the digitized data obtained from scans can be securely stored in digital formats, such as on computers, and easily transmitted via the Internet, facilitating rapid information exchange¹⁷. Consequently, this technology enables dental practitioners to make more accurate, rapid, convenient diagnoses, and administer treatments effectively. Because of it is efficiency and applicability IOS has been widely studied by scholars in the field of forensic odontology for bitemark analysis¹⁸, human identification based on palatal morphology¹⁹, and sex identification by analysing the morphology of palatal rugae²⁰. Furthermore, IOS demonstrates the same reliability, reproducibility, validity, and accuracy as cone-beam computed tomography in obtaining dental linear measurements for diagnostic purposes^21,22.

AI is another technological advancement that integrated deeply into day-to-day contemporary human life, with its presence extending across various facets from smartphones and automobiles to aviation, banking, healthcare, agriculture, scientific research, entertainment, and beyond. AI’s influence is pervasive, rendering it nearly omnipresent, thus modern forensic odontology is not an exception²³. ML, a subset of AI, focuses on the creation of algorithms and statistical models that empower computers to learn from data without direct programming. ML encompasses a range of applications such as detecting fraud, recommendation systems, and predictive modeling. In medical fields, the predominant approach employed by most ML algorithms is supervised learning. Initially, a preprocessing algorithm is employed to extract pertinent features from raw subject samples, thereby forming training and testing sets. Each dataset is structured as a matrix comprising features and records, Various classifiers are trained using the training set, and their efficacy is evaluated on the testing set. Clinically relevant classifiers are those demonstrating robust generalizability to unseen data²⁴.

In forensic applications, ML models demonstrate superior performance characterized by high prediction accuracy and relatively minimal bias when compared with traditional statistical tests²⁵. However, numerous studies have directed their attention toward conventional techniques for dental identification, while relatively few have explored ML methodologies and computer vision techniques for sex and age identification based on dental characteristics^26,27. Accordingly, this study aims to bring modern technology (3D images from IOS and ML) together and reevaluate Rao’s methodology to predict sex.

Results

One hundred ninety-six digital impressions were included in this study (98 males and 98 females). The average age for male subjects was 38.39 years with a standard deviation (SD) of 10.82, and for female subjects, it was 35.98 years with an SD of 11.14. There was no statistically significant difference in mean age between male and female samples (p = 0.126). Figure 1 shows the distribution of various morphometric values, including right and left mandibular canine measurements, MICD, and mandibular canine indices, for both male and female samples.

Fig. 1.

Scatterplot showing the distribution of the values of mandibular right canine height (MRCH), mandibular left canine height (MLCH), mean mandibular canine height (MeanMCH), mandibular right canine width (MRCW), mandibular left canine width (MLCW), MeanMCW, mandibular right canine ratio (MRCR), mandibular left canine ratio (MLCR), mean mandibular canine ratio (MeanMCR), MICD, mandibular right canine index (MRCI), mandibular left canine index (MLCI), and MCI in both sexes.

The results of the Shapiro–Wilk test indicated that the MRCW, MLCW, MeanMCW, MRCR, MLCR, MeanMCR, MRCI, MLCI, and MCI followed a normal distribution, whereas the MRCH, MLCH, MeanMCH, and MICD exhibited nonparametric distributions. The values of all parameters showed statistically significant differences between females and males, the mean ± SD of parametric data and the results of independent-T tests were shown in Table 1, while Table 2 indicates the outcome of the Mann–Whitney U test and descriptive statistics of nonparametric data.

Table 1.

Show the mean ± SD of the parametric data and the results of the independent-T tests.

Parameters	Sex	Mean ± SD	P value
MRCW (mm)	Female	5.773 ± 0.333	0.001
	Male	6.449 ± 0.395
MLCW (mm)	Female	5.766 ± 0.368	0.001
	Male	6.46 ± 0.468
MeanMCW (mm)	Female	5.769 ± 0.323	0.001
	Male	6.454 ± 0.409
MRCR	Female	0.734 ± 0.063	0.012
	Male	0.708 ± 0 0.078
MLCR	Female	0.727 ± 0.065	0.022
	Male	0.704 ± 0.075
MeanMCR	Female	0.73 ± 0.057	0.01
	Male	0.706 ± 0.072
MRCI	Female	0.2292 ± 0.0182	0.001
	Male	0.2469 ± 0.0179
MLCI	Female	0.2288 ± 0.0166	0.001
	Male	0.247 ± 0.0165
MCI	Female	0.229 ± 0.0165	0.001
	Male	0.247 ± 0.0164

Table 2.

Indicate the descriptive statistics of nonparametric values and the results of the Mann–Whitney U test.

Parameters	Sex	Median (Interquartile range)	P value
MRCH (mm)	Female	7.902 (0.92)	0.001
	Male	9.007 (1.49)
MLCH (mm)	Female	7.914 (0.79)	0.001
	Male	9.181 (1.69)
MeanMCH (mm)	Female	7.903 (0.87)	0.001
	Male	9.108 (1.59)
MICD (mm)	Female	25.355 (2.47)	0.001
	Male	26.208 (2.34)

The ROC analysis was performed to evaluate the ability of each parameter to predict sex. The ROC curves for all parameters are represented in Fig. 2. MeanMCW and MRCW demonstrated the highest AUC values (0.912 and 0.908, respectively), followed by MLCW (0.896) and MeanMCH (0.868). In contrast, MeanMCR exhibited the lowest discriminative power in predicting sex (AUC = 0.389) among all parameters. The ROC analysis showed statistical differences between all the parameters in males and females, with p < 0.05. Table 3 indicates the AUC, cut-off values, p-value, sensitivity, and specificity of all parameters.

Fig. 2.

Illustrates the ROC Curve of all parameters.

Table 3.

The results of ROC analysis.

Parameters	AUC (%95 CI)	Cut-off	P value	Sensitivity	Specificity
MRCH (mm)	0.860 (0.809–0.91)	8.3625	0.001	0.776	0.765
MLCH (mm)	0.862 (0.811–0.913)	8.3965	0.001	0.786	0.776
MeanMCH (mm)	0.868 (0.82–0.917)	8.3855	0.001	0.765	0.745
MRCW (mm)	0.908 (0.866–0.951)	6.0805	0.001	0.837	0.837
MLCW (mm)	0.896 (0.849–0.943)	6.0905	0.001	0.847	0.857
MeanMCW (mm)	0.912 (0.867–0.956)	6.0605	0.001	0.888	0.878
MRCR	0.407 (0.327–0.487)	0.7174	0.025	0.469	0.418
MLCR	0.398 (0.319–0.478)	0.7140	0.014	0.449	0.439
MeanMCR	0.389 (0.31–0.468)	0.7207	0.007	0.459	0.418
MICD (mm)	0.635 (0.558–0.712)	25.5850	0.001	0.612	0.582
MRCI	0.753 (0.686–0.821)	0.2376	0.001	0.724	0.673
MLCI	0.779 (0.715–0.843)	0.2384	0.001	0.704	0.745
MCI	0.779 (0.715–0.844)	0.2363	0.001	0.776	0.714

The logistic regression (LR) algorithm exhibited the greatest discriminative capability among all ML models with an AUC of 0.963. Following LR, GNB and random forest (RF) showed AUCs of 0.958 and 0.95, respectively, while the K-Nearest Neighbour (KNN) classifier achieved an AUC of 0.9. Conversely, the decision tree (DT) algorithm demonstrated the lowest AUC among all ML models with a value of 0.8. Figure 3 illustrates the ROC and AUC of all ML models.

Fig. 3.

Displays the ROC curve and AUC values for all ML algorithms.

The GNB algorithm exhibited the highest testing accuracy (0.925) and the lowest training accuracy (0.865) among all ML algorithms. Its confusion matrix correctly predicted 19 out of 20 females and 18 out of 20 males. Following GNB, The LR, KNN, and RF showed similar testing accuracy (0.90). Conversely, DT had the lowest testing accuracy (0.85) and highest training accuracy (1), correctly predicting 17 out of 20 females and males. The f1-scores for GNB were the highest for both females and males (0.927, 0.923 respectively), while LR, KNN, and RF followed closely with a score of 0.90 for both genders. In contrast, DT exhibited the lowest f1-score for both sexes (0.85). Table 4 summarizes the precision, recall, f1-score, training accuracy, and testing accuracy of all ML algorithms, whereas Fig. 4 illustrates the results of the confusion matrix of all ML models.

Table 4.

Shows the performance of all ML models.

ML algorithms	Sex	Precision	Recall	f1-score	Training accuracy	Testing accuracy
LR	Female	0.90	0.90	0.90	0.884	0.90
	Male	0.90	0.90	0.90
KNN	Female	0.90	0.90	0.90	0.89	0.90
	Male	0.90	0.90	0.90
GNB	Female	0.905	0.95	0.927	0.865	0.925
	Male	0.947	0.90	0.923
DT	Female	0.85	0.85	0.85	1	0.85
	Male	0.85	0.85	0.85
RF	Female	0.90	0.90	0.90	0.987	0.90
	Male	0.90	0.90	0.90

Fig. 4.

The confusion matrix of all ML models.

SHAP values have been employed to show the mean impact of each parameter on the RF and DT algorithms’ magnitude, the result of SHAP values are illustrated in Fig. 5.

Fig. 5.

The result of SHAP values: (A) DT, (B) RF.

In terms of the study’s reliability, the assessment of intra-examiner reliability for all parameters yielded intraclass correlation coefficients ranging from 0.773 to 0.975, Additionally, tenfold cross-validation was employed to evaluate the performance of the machine learning (ML) algorithms. The LR algorithm demonstrated the highest accuracy (0.878 ± 0.038), followed by GNB (0.87 ± 0.04) and KNN (0.857 ± 0.028), with DT exhibiting the lowest accuracy (0.833 ± 0.057). Table 5 presents the results of the tenfold cross-validation tests for all ML algorithms.

Table 5.

The results of tenfold cross-validation of all ML models.

Testing set		GNB		RF		DT		LR		KNN
1		0.85		0.80		0.85		0.85		0.825
2		0.875		0.90		0.90		0.925		0.90
3		0.80		0.75		0.775		0.825		0.825
4		0.95		0.875		0.80		0.95		0.85
5		0.825		0.825		0.70		0.825		0.825
6		0.875		0.85		0.875		0.90		0.85
7		0.90		0.875		0.85		0.875		0.85
8		0.875		0.85		0.825		0.875		0.90
9		0.85		0.875		0.875		0.875		0.89
10		0.90		0.875		0.875		0.875		0.85
Mean ± SD		0.87 ± 0.04		0.848 ± 0.043		0.833 ± 0.057		0.878 ± 0.038		0.857 ± 0.028

Discussion

In contemporary forensic investigations, sex determination has emerged as a crucial aspect and it has become a vital step in the process of identification of unknown human remains²⁸, for this purpose, odontometric analysis has been frequently used, and many studies concluded that the utilization of odontometric parameters presents a viable, cost-effective, and reliable source for identifying sex²⁹. Disagreement among scholars regarding the reliability of utilizing morphometric parameters of lower canines, MICD, and MCI for sex identification¹² prompts this study to reassess the value of canine odontometric parameters of mandibular canines, MCID, and MCI using innovative techniques like digital impressions and ML algorithms. Hence, this research aims to evaluate the accuracy of incorporating these new methodologies in sex estimation.

Two widely described and utilized methods for sex identification are metric and nonmetric. This study applied the metric approach which provides several advantages over the nonmetric method, being inherently more objective. It also boasts greater reliability, reduced reliance on the observer’s prior experience, and is more amenable to statistical analysis, which facilitates comparisons both within the sample and with previous studies^30,31. In this study, IOS has been used, unlike intraoral analysis digital impressions can be analysed at another time to compare the reliability of the measurements and to reduce errors during measurements due to fatigue³². Moreover, digital impressions are notable for their simplicity and the elimination of technical errors common in traditional impressions, such as dimensional changes in impression materials and air-bubble entrapment in the cast³³.

In this research, all odontometric parameters of the mandibular left and right canines (MRCH, MLCH, MeanMCH, MRCW, MLCW, MeanMCW, MRCR, MLCR, MeanMCR, MICD, MRCI, MLCI, and MCI) showed statistically significant differences between males and females p < 0.05. Males had higher mean values for all measured parameters than females expect of MRCR, MLCR, and MeanMCR. The results of the literature support the concept that morphometric variances in canine dimensions can manifest at the millimeter scale, with males typically exhibiting larger teeth compared to females^34–37. Nevertheless, it’s crucial to acknowledge that the reverse scenario is possible. Previous studies have documented instances where female subjects have been observed to possess larger teeth, including incisors, canines, and premolars, compared to male subjects, particularly within certain populations such as those in, India, and Odisha^38–40.

The results from ROC analyses across all parameters indicate that MeanMCW exhibits the highest sensitivity and specificity in sex prediction (AUC = 0.912) followed by MRCW and MLCW with AUC of 0.908, 0.896 respectively these results are supported by Azevedo et al.⁴¹ were they found that the results of AUC of the MRCW were 0.90 and 0.899 for the MLCW. Among the ML models evaluated, GNB demonstrates the highest accuracy in predicting sex at 92.5%, followed by KNN, LR, and RF with accuracies of 90%. Conversely, DT exhibits the lowest prediction accuracy at 85%. This study utilizes AUC and ROC to assess the diagnostic efficacy of various odontometric parameters such as mandibular canine dimensions, MICD, and mandibular canine indices in sex prediction, and the determination of optimal cut-off values for each predictor. The results of AUC have also been used to evaluate the capability of each classifier’s capacity to differentiate between males and females. Scholars have pointed out the benefits of employing the ROC curve and AUC, noting that the AUC considers both specificity and sensitivity, thus remaining uninfluenced by the prevalence of either group under investigation, unlike individual measures such as specificity, sensitivity, and diagnostic accuracy. Additionally, AUC and ROC curves offer a means to compare various models and are not impacted by class imbalances⁴².

Depending solely on mandibular canine indexes the overall accuracy in estimating sex by Iqbal et al.⁴³ was 76.85%, Patel et al.³² was 78.75%, Bakkannavar et al.⁴⁴ was 74.8%, and Kiran et al.⁴⁵ was 72%, while Acharya et al.⁴⁶ found that the accuracy of the MCI was 50.2%, the MD width of the mandibular right canine 65.7%, and the MICD 61.8%. The differences between the results of previously mentioned studies with this study could be attributed to the fact that in this study ML algorithms have been used dependent on the learning of all parameters used for training them in determining the accuracy⁴⁷ while other research that uses conventional statistics are depending on one parameter which is mainly the MCI. Unfortunately, no study has been found that has utilized ML classifiers to predict sex based on mandibular canine dimensions, MICD, and MCI to compare it with the result of this study. Azevedo et al.⁴¹ evaluated odontometric measurements of mandibular canines and MICD obtained from 120 plaster casts of both sexes and analysed the data by using ROC and LR analysis. Their findings are aligned with the findings of this research as the MD width of lower canines had higher sexual dimorphism among all other measurements and the overall sex determination of all parameters was 88.6%.

To evaluate the accuracy of repeatability of data recorded, intraexaminer reliability was performed and analysis by using the intraclass correlation coefficient, the results ranged from 0.773 to 0.975. Following the guideline for interpreting intraclass correlation coefficient outcomes, values falling within the range of 0.75 to 0.9 signify good reliability, while values greater than 0.90 represent excellent reliability⁴⁸. In this study, ten-fold cross-validation has been used which is a commonly used method to assess the effectiveness of a classification algorithm and to compare different algorithms on a dataset. In this method, the dataset is divided randomly into 10 separate and non-overlapping subsets of roughly equal size. Each subset is then used as a testing set while the model trained on the remaining 10 subsets is used for classification. This process is repeated 10 times, rotating which subset is used for testing each time. The performance of the classification algorithm is then measured by averaging the accuracies obtained across all 10 iterations³⁴. The advantages of using ten-fold cross-validation include the prevention of overfitting and bias. These advantages have made it widely used in machine learning methods for regression and classification⁴⁹.

This study is constrained by certain limitations inherent to its retrospective nature. As a result, it was unable to acquire full upper and lower digital impressions from all individuals included in the study, thereby hindering the assessment of orthodontic relationships between the upper and lower jaws and precluding the determination of whether subjects had undergone orthodontic interventions. Additionally, this study was confined to a particular geographic area with a limited sample size, and it is acknowledged that dental racial disparities may influence the outcomes, thus underscoring a further limitation.

Methods

Study design

The research spanned from November 2022 to March 2024. Ethical clearance for this study was acquired from the Ethics Committee at the College of Dentistry/University of Sulaimani under reference No. 36/21, granted on August 11, 2021, and all methodology of this study performed in accordance with the relevant guidelines and regulations of this above-mentioned committee. Informed consent requirements were waived by the Ethics Committee at the College of Dentistry/University of Sulaimani given that analysis primarily involved archival digital impression data.

Eight hundred and fifty-nine digital impressions of Kurdish subjects were collected from the B&R Dental Center archive in Sulaimani, Iraq, by a dental specialist with 10 years of experience in image analysis. The digital impressions including the mandibular dental arch from the 1st right mandibular premolar to the 1st left mandibular premolar, belonged to individuals aged 18 years and older of both sexes, were incorporated into the study. Conversely, digital impressions that contained the following criteria were excluded from consideration:

1. Mandibular anterior teeth that contained dental caries.

2. Any missing or extracted anterior lower teeth and/ or any prosthetic replacement of the anterior lower teeth.

3. Attrition of mandibular canine teeth reaching contact points.

4. Dental arches that contain filling or prosthodontic treatment of anterior teeth.

5. Any digital impressions with signs of orthodontic treatment.

6. Presence of spacing or crowding.

7. Lower anterior teeth with periodontal disease (root exposed reaching the middle third of the root or samples that have gingival hyperplasia).

8. Present of any deciduous teeth or restoration of lower anterior teeth.

As a result of the inclusion and exclusion criteria, 196 digital impressions were included in this study (98 males and 98 females), aged 18–67 years. All the digital impressions were acquired using the CEREC Omnicam IOS, manufactured by Dentsply Sirona, Charlotte, NC, USA. All digital impressions were taken by using CEREC 4.6.x Software developed by Dentsply Sirona, Charlotte, NC, USA.

Image analysis

The 3D images of digital impressions were exported in the stereolithography (STL) file format by using CEREC 4.6.x Software. The STL files were opened in the 3D Slicer version 4.13.0 software, an open-source software platform for biomedical research, and the following linear measurements were obtained by the same examiner that collected the digital impressions:

1. Right and left mandibular canine height (from the tip of the cusp until the lowest point of the cementoenamel junction).

2. MICD: the distance between the tips of the right and left mandibular canines.

3. Right and left mandibular canine width: the distance from the contact point of the lower lateral incisor to the contact point of the lower 1st premolar.

4. The formula that has been suggested by Rao et al.¹¹ was used to calculate the right and left mandibular canine indices:

• MRCI = MRCW/MICD

• MLCI = MLCW/MICD

• MCI = MeanMCW/MICD

MeanMCH, MeanMCW, MRCR, MLCR, MeanMCR, and the mandibular canine indices (MRCI, MLCI, and MCI) were calculated using Microsoft Excel 2021, developed by Microsoft Corporation, Redmond, WA, USA. The intra-examiner reliability test was conducted on 30 randomly collected samples, after six months. All odontometric measurements of mandibular canines and MICD are shown in Fig. 6.

Fig. 6.

Illustrate the odontometric measurements of mandibular canines and MICD. (A): MICD, (B): MRCW, (C): MRCH, (D): MLCW, (E): MLCH.

Statistical analysis

Statistical Package for the Social Sciences (SPSS) Version 25, developed by International Business Machines Corporation (IBM), New York, NY, USA was used to perform all statistical tests. The mean ± SD, median, and interquartile range were calculated by employing descriptive statistics, whereas the Shapiro–Wilk normality test was applied to assess the normality of the data.

To analyse the differences between male and female parametric data, the independent t-test was utilized, while the nonparametric data were asses by employing the Mann–Whitney U test. The intra-examiner reliability was evaluated by using the intraclass correlation coefficient.

The ROC has been used to report the discriminative power of each parameter in predicting the sex. All results were considered statistically significant at less than p < 0.05.

ML algorithms

In this study, ML models were created using Anaconda Navigator version 2.1.4 and Jupyter Notebook version 6.4.8, both developed by Anaconda Inc., based in Austin, TX, USA. These tools were utilized within the Python programming language version 3.10, developed by The Python Software Foundation, located in Wilmington, DE, USA. The ML modelling process was executed on MSI personal computer Model: GF65, featuring a 9th generation Intel Core i7 processor. Various algorithms including GNB, LR, DT, RF, and KNN were employed for data analysis. The dataset was divided, allocating 80% for training and 20% for testing purposes. Additionally, to evaluate the reliability of the ML models, tenfold cross-validation was conducted by shuffling the dataset and running the ML algorithms ten times. The mean and SD of each algorithm’s accuracy were calculated using Microsoft Excel 2021, developed by Microsoft Corporation, headquartered in Redmond, WA, USA. Figure 7 illustrates the summary of the methods that were used in this research.

Fig. 7.

A diagram summarizing the methodology of this study.

Author contributions

M.T.A.B.: conceptualization, methodology, formal analysis, data curation, software, writing—original draft; D.N.M.: supervision, project administration, resources, writing—review and editing.

Funding

This research received no external funding.

Data availability

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

Declarations

Competing interests

The authors declare no competing interests.