Artificial intelligence-based prescription of personalized scalp cosmetics improved the scalp condition: efficacy results from 100 participants

Abstract

Background: Scalp-related symptoms such as dandruff and itching are common with diverse underlying etiologies. We previously proposed a novel classification and scoring system for scalp conditions, called the scalp photographic index (SPI); it grades five scalp features using trichoscopic images with good reliability. However, it requires trained evaluators.Aim: To develop artificial intelligence (AI) algorithms for assessment of scalp conditions and to assess the feasibility of AI-based recommendations on personalized scalp cosmetics.Methods: Using EfficientNet, convolutional neural network (CNN) models (SPI-AI) ofeach scalp feature were established. 101,027 magnified scalp images graded according to the SPI scoring were used for training, validation, and testing the model Adults with scalp discomfort were prescribed shampoos and scalp serums personalized according to their SPI-AI-defined scalp types. Using the SPI, the scalp conditions were evaluated at baseline and at weeks 4, 8, and 12 of treatment.Results: The accuracies of the SPI-AI for dryness, oiliness, erythema, folliculitis, and dandruff were 91.3%, 90.5%, 89.6%, 87.3%, and 95.2%, respectively. Overall, 100 individuals completed the 4-week study; 43 of these participated in an extension study until week 12. The total SPI score decreased from 32.70 ± 7.40 at baseline to 15.97 ± 4.68 at week 4 (p < 0.001). The efficacy was maintained throughout 12 weeks.Conclusions: SPI-AI accurately assessed the scalp condition. AI-based prescription of tailored scalp cosmetics could significantly improve scalp health.

Keywords:

• Artificial intelligence
• dandruff
• evaluation
• folliculitis
• prescription
• scalp

Introduction

Scalp-related symptoms, such as dandruff and pruritus, are considerably common. A dandruff survey conducted in China and the United States revealed dandruff prevalence rates of 30–40%, 66–82%, and as high as 81–95% among the Chinese, Caucasian, and African-American populations (1(1) $$\text{Accuracy}=\left(\text{true positive}+\text{true negative}\right)/\left(\text{positive}+\text{negative}\right)$$(1) ). Furthermore, the prevalence of scalp pruritus is reported to be 13–45% (2). These symptoms are caused by various underlying scalp conditions, which must be assessed accurately to enable personalized treatment. However, no systematic classification systems have been established for the scalp to describe diverse scalp characteristics and objectively quantify the overall conditions, unlike for glabrous skin (for which classification systems such as the Baumann skin type are widely used) (3). One of the reasons is the inherent difficulty in directly observing the scalp. Several studies employed dermoscopy to observe the scalp structures in detail and designed various scalp assessment systems for their outcome measures. However, these approaches failed to provide a comprehensive evaluation and lacked validation (4–8).

Therefore, we previously devised the scalp photographic index (SPI) to address this issue (9). The SPI is an instrument that classifies the scalp types and scores the severity of the following five features of scalp conditions: (1(1) $$\text{Accuracy}=\left(\text{true positive}+\text{true negative}\right)/\left(\text{positive}+\text{negative}\right)$$(1) ) dryness: fine lines or tiny scales on the scalp surface, (2) oiliness: presence of greasy sebum or shine on the scalp surface, (3) erythema: erythema or telangiectasia between the hair pores, 4) folliculitis: erythema or pustules of the hair pores, and (5) dandruff: coarse and large white scales on the scalp surface. Based on magnified (× 60) scalp photographs and a standard photographic scale, the severity of these five features is graded as follows: 0 (none), 1 (mild), 2 (moderate), and 3 (severe). The sum of each score reflects the overall severity of the scalp problems and a feature of grade ≥2 enables the classification of the scalp types. We previously demonstrated that the SPI is correlated with the individual’s symptoms and the investigator’s visual evaluation. Furthermore, the SPI has excellent internal consistency and inter- and intra-rater reliability.

Using this system, physicians can provide treatments suitable for the specific scalp type and monitor the treatment response using the total SPI score. However, the wider application of the SPI is limited by the fact that grading requires trained evaluators. Automating the evaluation will enable a rapid and reliable assessment of scalp conditions; this will benefit not only healthcare providers, but also the beauty industry and general population. Furthermore, the automated system would also facilitate customization of scalp cosmetics.

In this study, we aimed to establish artificial intelligence (AI) algorithms to assess scalp conditions and evaluate whether they can be improved by an AI-based personalized prescription of scalp cosmetics.

Materials and methods

Development of AI algorithms for scalp photographic index scoring (SPI-AI)

Dataset

A total of 163,870 60-fold magnified images were acquired from 32,774 Koreans using a trichoscope (ASM-202^TM camera, Aram Huvis Co., Ltd, Seongnam, Korea), and 126,948 images, excluding those where the scalp was obscured by hair, were included in the dataset. The demographics are summarized in Table 1. All images were uploaded to a web-based annotation program developed by HealthCloud Co., Ltd, (Seoul, Korea) for data cleaning, labeling, and review.

Table 1. Demographics of the 32,774 Koreans whose scalp images were used to develop the SPI-AI.

Age (years)	N (men:women)
0–9	345 (125:220)
10–19	1,057 (477:580)
20–29	8,172 (3,161:5,011)
30–39	8,307 (4,527:3,780)
40–49	6,266 (2,863:3,403)
50–59	5,180 (1,855:3,325)
60–69	2,645 (668:1,977)
70–79	641 (151:490)
80–89	146 (46:100)
90–99	14 (5:9)
≥100	3 (3:0)
Total	32,774 (13,879:18,895)

SPI-AI: scalp photographic index-artificial intelligence

Step 1: Data cleaning

After removing duplicates and poor-quality images, a total of 101,027 images were selected.

Step 2: Data labeling (SPI) and review

For each image, SPI grading was initially performed by trained laypersons, and then by two experts (J. W. S. and H. R. C.) who subsequently reviewed the grades and one expert (J. I. N.) who confirmed the final grades.

Step 3: Training, validation, and test sets

Images for each scalp feature were divided into the training, validation, and test sets in a 7:2:1 ratio.

Algorithm training

Using EfficientNet-B7 which achieved the highest accuracy on the ImageNet classification task in 2019 and surpassed other convolutional neural networks (CNNs) (10), we developed five CNN models SPI-AI that can classify each of the five scalp features into grades 0–3. We fine-tuned EfficientNet using PyTorch 1.7 for 50 epochs with a batch size of 128 (width coefficient 2.0; depth coefficient 3.1; default resolution 600; dropout rate 0.5). All the layers were frozen except the final fully connected layer.

Accuracy assessment

The accuracies of the five CNN models were assessed using the following equation: (1) $$\text{Accuracy}=\left(\text{true positive}+\text{true negative}\right)/\left(\text{positive}+\text{negative}\right)$$(1)

For example, if the model predicted a grade 3 image as being a grade 3 image, the prediction was deemed correct. Conversely, if the model predicted a grade 3 image as being a grade 1 or 2 image, the prediction was deemed incorrect.

Study design and population

This prospective, single-center, pre–post comparative study was approved by the institutional review board (IRB No. B-2107-697-301) and was performed in accordance with the Declaration of Helsinki. All participants provided written informed consent prior to participation. The inclusion criteria were as follows: (1(1) $$\text{Accuracy}=\left(\text{true positive}+\text{true negative}\right)/\left(\text{positive}+\text{negative}\right)$$(1) ) healthy adults aged >19 years and (2) presence of scalp-related discomfort (such as itching, a sensation of heat, dandruff, excessive sebum, dryness, pain, or pimple). The exclusion criteria were as follows: (1(1) $$\text{Accuracy}=\left(\text{true positive}+\text{true negative}\right)/\left(\text{positive}+\text{negative}\right)$$(1) ) presence of active infectious scalp diseases and (2) history of treatment with topical or systemic steroids, oral immunosuppressants, topical or systemic antibiotics, or antifungal agents that could affect the scalp conditions, within 4 weeks prior to the baseline. Furthermore, participants who had difficulty understanding or following the protocol were also excluded at the discretion of the investigator. In total, 100 participants met the eligibility criteria and were enrolled in this study. All participants had not been included in the previous image acquisition for training of SPI-AI.

For each participant, 60-fold magnified images were acquired from five standard areas (vertex, anterior, posterior, right, and left) of the scalp using a trichoscope; following an SPI-AI analysis, the scalp type was determined. A scalp feature of grade ≥2 was regarded as positive since it indicated clinically significant severity that needed treatment. Taken prescriptions for positive scalp features into account, the scalp types were classified into nine categories (normal, dry, oily, erythema, dry erythema, oily erythema, dry dandruff, oily dandruff, and folliculitis) as shown in Figure 1.

Figure 1. Classification of scalp types. scalp photographic index-artificial intelligence (SPI-AI) diagnosed a feature of grade ≥2 as positive and classified the scalp type according to the positivity. ^aIf both dryness and oiliness were positive in one scalp area, only oiliness of grade 3 was considered positive. For example, a scalp with grade 2 or 3 dryness, grade 2 oiliness, grade 2 or 3 erythema could be defined as a dry erythema type.

In cases when the scalp types of the five standard areas did not match each other, it was determined by setting priorities in the following order: folliculitis, oily erythema, dry erythema, oily dandruff, dry dandruff, erythema, oiliness, dryness, and normal.

For example, a scalp with grade 2 or 3 oiliness and dandruff could be defined as an oily dandruff type, whereas a scalp with grade 2 or 3 oiliness and folliculitis could be classified as a folliculitis type. If one area is classified as folliculitis and another area as oily erythema, a prescription for folliculitis is given. This is because the most numerous active ingredients are included in prescriptions for folliculitis.

Treatment

The personalized scalp cosmetics (shampoos and scalp serums) were manufactured by Aram Huvis Co. Ltd (Seongnam, Korea). The combination and concentrations of the ingredients in the shampoos and scalp serums were determined according to the patients’ scalp types classified by the SPI-AI analysis. Active ingredients included hyaluronic acid, dexpanthenol, inositol, piroctone olamine, allantoin, beta-glucan, and ceramide (11–20).

Study visits occurred at the baseline and week 4; during these, compliance and history of adverse events were examined under investigator supervision. On day 1, all participants were dispensed the customized shampoo and scalp serum and instructed on their usage as below.

1. Soak your hair with lukewarm water before applying shampoo.

2. Start lathering up from the roots, and then massage down toward the tips.

3. After scrubbing the scalp gently, rinse the shampoo out thoroughly for a minimum of three minutes.

4. After shampooing, dry the scalp with a towel and apply the scalp serum gently to the entire scalp using fingers.

Thereafter, the agents were applied once daily for 4 weeks at home. The study was extended until week 12 for volunteers.

Participants who used other topical or systemic treatments concomitantly for scalp conditions during the trial period, those who applied the study medications for less than 8 d out of 4 weeks, and those who experienced adverse events and could not continue the trial were regarded as drop-outs at the discretion of the investigator. This clinical study was conducted from 27 July 2021 to 7 January 2022.

Outcomes

A single dermatologist (J.I.N.) performed the SPI evaluation at every visit using the trichoscopic images of the participants’ scalp. The images were obtained using the same device with identical settings at each visit, and the scalp positioning ruler was used to replicate the exact spot of each image (Supplementary Figure 1). The participants were instructed against washing their scalps or using the scalp serum within 12 h before the evaluation. The primary endpoint was the change from the baseline (CFB) in the total SPI score at week 4; the key secondary endpoints were the CFB in the total SPI score at weeks 8 and 12.

A seven-item questionnaire was used to assess the subjective improvement in scalp conditions (namely, dryness, oiliness, dandruff, itching, scalp pimple, pain, and heating) relative to the baseline. The participant’s satisfaction was evaluated as follows: 0 (unsatisfied), 1 (neutral), 2 (satisfied), and 3 (very satisfied). The secondary endpoints comprised the proportion of participants achieving improvement in more than one item on the questionnaire and the proportion of participants satisfied with the products at week 4. Furthermore, the participants were asked to report on any adverse events that occurred at any time during the treatment. The investigators also checked for adverse events during each follow-up visit.

Statistical analysis

Each measurement at the follow-up visit was compared with the baseline value using paired t-tests. The results are expressed as mean ± standard deviation, and a p value of <0.05 was considered statistically significant. IBM SPSS version 20 (IBM Corp., Armonk, NY) was used to analyze all data.

Results

Accuracy of SPI-AI

Table 2 displays the SPI results of the 101,027 scalp images considered for algorithm training, validation, and testing. Dryness, oiliness, erythema, folliculitis, and dandruff were observed in 17,434, 80,416, 67,414, 4,592, and 40,482 of these images, respectively. A total of 811 images were graded 0 for all five features, implying good scalp conditions.

Table 2. Classification of 101,027 scalp images labeled with the severity grade based on the SPI scoring system.

			Severity
		0	1 (mild)	2 (moderate)	3 (severe)	Total
Scalp features	Dryness	4027 (18.76%)	6335 (29.52%)	7837 (36.52%)	3262 (15.2%)	21,461
	Oiliness	6080 (7.03%)	40,087 (46.34%)	34,978 (40.44%)	5351 (6.19%)	86,496
	Erythema	6871 (9.25%)	42,799 (57.61%)	18,509 (24.92%)	6,106 (8.22%)	74,285
	Folliculitis	1239 (21.25%)	3036 (52.07%)	1082 (18.55%)	474 (8.13%)	5831
	Dandruff	3999 (8.99%)	23,656 (53.18%)	13,604 (30.58%)	3222 (7.24%)	44,481
	Normal		–	–	–	811

Multiple features could be observed in one scalp image. When severity 0 was confirmed in all five items of dryness, oiliness, erythema, folliculitis, and dandruff, it was classified as normal scalp.

SPI: scalp photographic index

Five CNN models, i.e., those for dryness, oiliness, erythema, folliculitis, and dandruff, were created. The training and test sets for each model and their accuracies are summarized in Table 3. Every model exhibited an accuracy of >85%; the dandruff model exhibited the highest accuracy (95.2%), followed by the dryness model (91.3%). Conversely, the folliculitis model exhibited the lowest accuracy (87.3%), possibly due to its smaller training set. When grades 0 and 1 were regarded as negative and grades 2 and 3 as positive, the sensitivity, specificity, and F1 score of each model were as follows: dryness (sensitivity 0.983, specificity 0.984, F1 score 0.984), oiliness 9, sensitivity 0.949, specificity 0.947, F1 score 0.944), erythema (sensitivity 0.839, specificity 0.999, F1 score 0.912), folliculitis (sensitivity 0.794, specificity 1.000, F1 score 0.885), dandruff (sensitivity 0.980, specificity 0.968, F1 score 0.964) (Supplementary Figure 2).

Table 3. Accuracy of the SPI-AI.

	Training set (n)	Validation set (n)	Test set (n)	Accuracy (%)
Dryness model	15,023	4291	2147	91.3
Oiliness model	60,572	17,270	8654	90.5
Erythema model	52,001	14,856	7428	89.6
Folliculitis model	4,082	1165	584	87.3
Dandruff model	31,137	8896	4448	95.2

SPI-AI: scalp photographic index-artificial intelligence

Participant characteristics

All the 100 participants enrolled in the study completed the treatment. The mean age was 34.90 ± 7.85 years (range, 19–61 years), and women accounted for 69% of the study population. The mean total SPI score at the baseline was 32.70 ± 7.40 (range, 15–50). The scalp types were defined by SPI-AI, and were as follows: folliculitis, 73%; oily dandruff, 12%; oily erythema, 7%; dry erythema, 4%; dry dandruff, 3%; and erythema, 1% (Supplementary Table 1). Forty-three participants (mean age, 36.40 ± 8.34 years) further participated in the extension study and completed the schedule until week 12.

Primary and secondary endpoints

The CFB in the total SPI score was from 32.70 ± 7.40 to 15.97 ± 4.68 at week 4 (p < 0.001), and all scalp subtypes showed consistent results (Table 4). Participants enrolled in the extension study also demonstrated a significant decrease in the total SPI score. The CFB in the total SPI score was from 31.42 ± 8.33 at baseline to 15.33 ± 4.65 and 16.05 ± 4.80 at weeks 8 and 12, respectively (p < 0.001 for both; Table 5). Figure 2 shows the photographs of two representative cases that presented a marked improvement at week 4.

Figure 2. Sixty-fold magnified scalp photographs of two representative cases with significant improvement in the scalp photographic index (SPI) from before to after treatment. (a) S10 (29-year-old man, folliculitis type): the total SPI score decreased from 38 at the baseline to 14 at week 4. (b) S31 (26-year-old man, oily erythema type): the total SPI score decreased from 33 at the baseline to 16 at week 4.

Table 4. Primary outcomes after 4 weeks of treatment.

	Total SPI score (mean ± SD)
	Week 0	Week 4	p Value
Total (n = 100)	32.70 ± 7.40	15.97 ± 4.68	<0.0001
Scalp type
Folliculitis (n = 73)	34.99 ± 6.21	16.44 ± 4.58
Oily dandruff (n = 12)	23.75 ± 5.79	12.75 ± 4.69
Oily erythema (n = 7)	28.00 ± 2.65	17.29 ± 3.45
Dry erythema (n = 4)	36.50 ± 4.20	17.75 ± 5.56
Dry dandruff(n = 3)	24.00 ± 8.72	13.67 ± 4.62
Erythema (n = 1)	17	11

SD: standard deviation; SPI: scalp photographic index

Table 5. Changes from the baseline in the total SPI score at weeks 4, 8, and 12 in patients who participated in the extension study.

	Total SPI score (mean ± SD, n = 43)	p Value
Week 0	31.42 ± 8.33
Week 4	15.07 ± 4.74	<0.0001
Week 8	15.33 ± 4.65	<0.0001
Week 12	16.05 ± 4.80	<0.0001

SPI: scalp photographic index

At week 4, 93% of the participants achieved an improvement in more than one items on the questionnaire. Furthermore, 63% and 1% of the participants rated the customized products as satisfactory and unsatisfactory, respectively; 31% of the participants had a neutral opinion about these products.

Adverse events

Six of the 100 participants (6.0%) enrolled experienced aggravation of symptoms, such as itching, scalp pimples, pain, and a sensation of heat. All complications were mild, and discontinuation of the study and rescue treatment was not required. Four of the aforementioned six participants recorded improvements in symptoms other than the aggravated item on the questionnaire; five of these participants further evaluated the customized products as satisfactory. No other local adverse events occurred in the treatment area.

Discussion

Although scalp-related symptoms are considerably common, studies on their etiopathogenesis and treatment are limited (2, 21). Dandruff, one of the most common complaints of patients visiting dermatological clinics, is often considered an inflammatory condition related to Malassezia spp. (sebum-dependent yeasts); accordingly, antifungal agents are typically administered to patients (22). However, in a Korean study, approximately 46% of the patients with dandruff had dry scalps, with sebum levels lower than those of normal controls; this suggested differences in the etiopathogenesis between dry and oily dandruff (23). Many patients choose cosmetic products that are inappropriate for their scalp condition.

One of the reasons why scalp conditions have not been widely studied is that they are difficult to observe (8). People cannot directly observe their own scalps, and photographing the scalp is difficult due to hair coverage. Magnified scalp photographs facilitate an easier analysis of scalp conditions covered with hair; this not only allows the observation of both the hair follicles and inter-follicular scalp, but also renders it possible to determine whether the scalp problem has affected the follicle or the inter-follicular epithelium. We used a trichoscope, a dedicated camera for hair and scalp imaging, to capture the scalp images in our study. Unlike dermoscopes, trichoscopes are fixed-focusing, allowing non-experts to easily capture well-focused photos. They are cheaper than dermoscopes; therefore, they are potentially more affordable to both healthcare providers and the general population. The use of a dedicated camera for the scalp also enables the capture of standardized photographs, thereby facilitating the automation of image evaluation.

In this study, over 100,000 scalp images were collected and graded according to each feature’s severity. Through this unprecedented effort, it was possible to train an AI algorithm (SPI-AI) to automatically evaluates scalp conditions; this algorithm exhibited an excellent performance, despite using a simple training technique. The SPI-AI could classify scalp conditions by detecting the presence or absence of five common scalp symptoms (dryness, oiliness, erythema, folliculitis, and dandruff) from magnified scalp images; this facilitated the identification of an individual’s scalp type, which can be used for recommending a suitable product and administering tailored treatment for improved scalp health.

Although some researchers have proposed a deep learning-based inspection of the scalp (8), studies on the feasibility of AI usage in clinical care settings have been sparse to date. In this study, we demonstrated a significant improvement in the condition of all scalp subtypes after prescribing personalized care according to the algorithm’s classification of the scalp types. Subjective evaluations revealed high participant satisfaction. The participant satisfaction and high treatment adherence were in part attributable to standardized scalp photography, which provided objectivity and consistency to the evaluations (24, 25).

Our study has some limitations. These include non-comparability and a lack of biophysical measurements. Further randomized studies with pre- and post-treatment biophysical measurements and a long-term follow-up will help establish the validity of our findings. Given that this study focused on Koreans, studies with larger populations with different skin types are required to obtain generalizable results and advance the performance of the AI-based model.

In conclusion, we developed a deep-learning algorithm trained using a substantial number of magnified scalp images and confirmed its ability to both accurately grade the severity of five scalp-condition features and diagnose the scalp type. Our data suggest that SPI-AI potentially provides an automated, prompt, and accurate diagnosis of an individual’s scalp condition; thus, this tool may be useful in suggesting personalized and more-tailored treatments for improved scalp health. We foresee a widespread use of this tool by the hairdressing industry and the general public owing to its automated nature.

Acknowledgments

IRB approval status: reviewed and approved by the Institutional Review Board of Seoul National University Bundang Hospital (IRB No. B-2107-697-301).

Disclosure statement

In accordance with Taylor & Francis policy and my ethical obligation as a researcher, I am reporting that Aram Huvis Co., Ltd (Seongnam, Korea) provided a trichoscope (ASM-202TM) and manufactured the personalized shampoo and scalp serum that were all used in this study. I have disclosed these interests fully to Taylor & Francis, and have in place an approved plan for managing any potential conflicts arising from this arrangement.

Data availability statement

For any purpose, with investigator support, anyone can request the data through email-id [email protected]

Additional information

Funding

This work was supported by the Ministry of SMEs and Startups and the National Information Society Agency of Korea under Grant number 20200804-TM50.