Skin and subcutaneous conditions are a major, continuous load on health systems. Global Burden of Disease data show that fungal infections, acne and other skin and subcutaneous illnesses abound, but the distribution varies considerably from country to country in terms of disease incidence and access to specialists [1]. Common inflammatory diseases also have a substantial population impact. For instance, in the world over, about 204 million people suffer from atopic dermatitis, of whom more than 100 million are adults and more than 100 million are children [2]. In India, outpatient dermatology practice must deal with various infectious, inflammatory, pigmentary, appendageal and drug-induced disorders, often in situations where access to specialists and diagnostic services is uneven. The above-mentioned variety creates a need for correct first-contact diagnosis, but complicates the standardisation of diagnosis. AI-driven image analysis has come out as an influential decision-making technique for the above mentioned medical facilities. Recent investigations have shown variable performance in AI with skin disease tasks as the end result of different factors such as diversity of participants in studies, varying types of image sources, diseases characteristics and methods of evaluation [3, 4]. The outcome of a meta-analysis conducted for skin lesion analysis algorithms in 2024 showed the sensitivity and specificity results of the algorithms amounted to 86.3 and 78.4 as compared to experts in corresponding studies [5]. While those figures appear promising, they may not imply that the algorithms can be widely used in the ordinary healthcare practice. At present most evaluations are retrospective and only a limited number of types of diseases are studied, expert images are applied in experiments, etc., which makes the results hard to extrapolate to the real-life data [3-5]. The importance of external validation as a means of establishing transportability cannot be exaggerated. The International Skin Imaging Collaboration challenge has demonstrated how significantly user performance can be diminished by distribution changes, presence of artifacts in patient images, or absence of certain diagnoses in training data [6]. Moreover, the problem is aggravated by issues of equity: dermatology AI and even clinician-labeled datasets demonstrate poor effectiveness with respect to clothing or diseases related to darker-skinned patients and non-prevalent diseases [7]. The scoping review also discussed the absence of skin types in studies utilizing machine learning [8]. It is safe to conclude that a model which relies mostly on the images from the datasets will acquire some shallow rules that may not work in the real world. The diagnosis is not determined purely based on the imaging. Age, lesion duration, clinical signs, number of episodes, medical history, and previous medications should be taken into account when making the diagnosis. Some ideas about the integration of the information listed above have been developed recently. Some scholars introduced the idea of using multimodal AI, which provides assistance through the problems associated with the classical classification of images. Some special examples of various illnesses situated in developing countries indicate that it is possible to use the combination of clinical signs and rash images for diagnostic purposes. However, this application is not quite common in practical medicine. One of the novel technologies called foundation models of big language models provides a big potential of using images and text in the process. However, still, research has to be conducted in isolated places. The study aims to verify the accuracy of a previously developed, locked multimodal deep-learning model with the help of standard clinical photographs along with the structured history made during routine outpatient assessments. It was postulated that the model would show diagnostic concordance with a dermatologist followed diagnosis in a diverse real-world sample. The main aim was to find the top-1 accuracy. Other aims include calculation of top-3 accuracy, agreement, prediction confidence, diagnostic performance, and consistency evaluation.

Study design By design this was a one-centre observational external-validation study. Its objective was to validate an already existing multimodal AI model in an entirely separate clinical cohort and, during the study, the model was neither retrained nor fine-tuned or modified in any way. The manuscript was written in accordance with the standards of transparent AI prediction-model reporting, including a description of the validation cohort, reference standard, endpoints and uncertainty measures [15]. The available source documents did not provide the model architecture, training dataset, software version or prespecified model threshold which precludes retrospective reconstruction and is specifically identified as a limitation of reporting. Setting and study period The research was carried out in the Outpatient Department of Dermatology at Basaveshwara Teaching and General Hospital which is affiliated to Mahadevappa Rampure Medical College, Kalaburagi, Karnataka, India. Recruitment and assessment took place between 1st August 2025 and 1st February 2026. This hospital provides tertiary care to a wide variety of patients which is useful in the research process in terms of diagnosis. Study population This was a one-centre observational external-validation study. The study was conducyted to validate an already existing multimodal AI model in an entirely separate clinical cohort and, during the study, the model was neither retrained nor fine-tuned or modified in any way. The manuscript was written in accordance with the standards of transparent AI prediction-model reporting, including a description of the validation cohort, reference standard, endpoints and uncertainty measures [15]. The available source documents did not provide the model architecture, training dataset, software version or prespecified model threshold which precludes retrospective reconstruction and is specifically identified as a limitation of reporting. Patients were eligible if they presented with a dermatologic condition during the study period, had a standardised clinical photograph available for assessment, had structured clinical history documented in the study record, had a final confirmed diagnosis established by a board-certified dermatologist, and provided written informed consent for participation and image use. Patients were excluded if they declined to provide written informed consent, had incomplete clinical metadata, had photographs judged to be of poor quality and therefore unsuitable for model interpretation, or had an uncertain final diagnosis. Sample-size determination The planned sample size was 70. It was calculated using an a priori matched-pairs Wilcoxon signed-rank test under the A.R.E. method, assuming a normal parent distribution, two-sided alpha of 0.05, effect size dz of 0.44 based on Gaurav et al., and 95% power. The calculation yielded a required total sample size of 70, with an actual power of 0.9517. Clinical assessment, image acquisition and index test Once informed consent had been given in writing, the treating clinician noted the clinical background, symptoms, characteristics of the lesions and made a preliminary diagnosis. Photos taken using the usual protocol of non-dermoscopic photography were obtained and the recorded images were processed for similarity in lighting, background and distance. After being anonymized, the images were assigned a study number for the purposes of analysis. The structured clinical history, together with its corresponding photo, were introduced into the pre-designed multimodal model. This model was applied as a fixed index test, and the obtained top diagnosis ranking and its accompanying confidence score were determined independently. Also, the images and data from the subjects were never reused. statistical analysis The reference point for this study was the final dermatological diagnosis entered into the clinical record. Top-1 accuracy was chosen as the main outcome of the study which refers to the case when the doctor’s diagnosis matches the prediction by the machine learning model. Secondary endpoints included Top-3 accuracy, Cohen kappa agreement, confidence score distinction between correct and incorrect top-1 predictions, sensitivity and specificity, and subgroup accuracy by sex, age, quality level, and number of injuries, etc. Results shown in the descriptive data are presented in terms of SD (standard deviation). The referred study calculated 95% CI (confidence interval) for accuracy estimates, chi-square comparisons for categorical variables and t-tests for the comparison of confidence scores. Two-sided p-values lower than 0.05 are considered statistically significant, due to the fact that counts of category-specific variables are low Ethical Consideration Prior to the gathering of history and photographic evidence, written consent was taken. No experimental treatment was done nor was any invasive procedure carried out. The images were anonymised and all the study material was recorded using unique numbers. In the source document, it is mentioned that the study qualified for its instititional ethics commitment and it was conducted according to Helsinki Declaration and ethical guidelines of Indian Council of Medical Research. It is to be pointed out that the ethics approval identifier was not provided and should come to the fore before submitting this paper to the journal.

Participant characteristics and clinical spectrum A total of 70 subjects were included in the study. The average age was 40.06 years (SD 18.60), and out of 70 subjects, there were 44 women (62.9% of the total). The age group of 18 to 59 years comprised 75.7% of the study group subjects. The most frequent symptom of presentation was itching (37.1%), followed by asymptomatic presentation (31.4%), and 22.9% of the patients reported a prior illness. Multiple lesions were recorded in 51.4% of the patients. Among the diagnosed illnesses, the commonest was dermatophytosis (17.1% of patients), while the second most common was acne vulgaris (12.9%) succeeded by atopic dermatitis/eczema (11.4%), onychomycosis (8.6%) and pityriasis versicolor (7.1%) (Table 1). Table 1. Participant characteristics and clinical spectrum (N=70) Characteristic n (%) or mean +/- SD Age, years 40.06 +/- 18.60 Age <18 years 8 (11.4) Age 18-39 years 26 (37.1) Age 40-59 years 27 (38.6) Age >=60 years 9 (12.9) Female sex 44 (62.9) Male sex 26 (37.1) Itching 26 (37.1) Asymptomatic 22 (31.4) Previous similar illness 16 (22.9) Multiple lesions 36 (51.4) Most frequent final diagnoses Dermatophytosis 12 (17.1); acne vulgaris 9 (12.9); atopic dermatitis/eczema 8 (11.4); onychomycosis 6 (8.6); pityriasis versicolor 5 (7.1) SD, standard deviation. Percentages may not total 100 because symptom categories were recorded as observed combinations in the source dataset. Image quality and model confidence The majority of the photos received good quality ratings (58/70, 82.9%), while the remaining photos received sufficient quality ratings (12/70, 17.1%); none of the analyzed photographs were considered to be of low quality. The median score for the best initial prediction was 0.76 (interquartile range: 0.68-0.87; observed range: 0.42-0.95). The confidence proved to be remarkably bigger in case the top-ranked model presented a correct diagnosis versus an incorrect one (mean 0.79 [SD 0.10] vs. 0.61 [SD 0.13]; t=5.96, p<0.001) (see Table 2). Table 2. Image quality and top-1 model confidence score (N=70) Parameter Result Image quality: good 58 (82.9%) Image quality: adequate 12 (17.1%) Top-1 confidence, mean +/- SD 0.75 +/- 0.13 Top-1 confidence, median (IQR) 0.76 (0.68-0.87) Observed confidence range 0.42-0.95 Concordant predictions (n=55) 0.79 +/- 0.10 Discordant predictions (n=15) 0.61 +/- 0.13 Between-group comparison t=5.96; p<0.001 IQR, interquartile range; SD, standard deviation. Concordance refers to agreement between the AI top-1 output and final confirmed diagnosis. Primary diagnostic accuracy and agreement The main goal of the study, which is to achieve top-1 accuracy, was 78.6% (55/70; 95% CI 67.6%-86.6%). If we take into account the top three diagnoses suggested by the model, the accuracy would rise up to 92.9% (65/70; 95% CI 84.3%-96.9%). There was a high level of agreement between AI top-1 output and the final diagnosis (Cohen’s kappa 0.767), and there was an almost perfect agreement between the dermatologist’s initial diagnosis and the final diagnosis (kappa 0.814). The association between observed top-1 and top-3 values is valid in terms of chi-square analysis (p<0.001). The ability of the score to distinguish correct from incorrect top-1 predictions in terms of AUC is quite good (AUC 0.866) (Table 3). Table 3. Overall diagnostic performance and agreement with final confirmed diagnosis Metric Value 95% CI or analysis Top-1 accuracy 78.6% (55/70) 67.6%-86.6%; chi-square=22.86, p<0.001 Top-3 accuracy 92.9% (65/70) 84.3%-96.9%; chi-square=51.43, p<0.001 AI top-1 vs final diagnosis, kappa 0.767 Substantial agreement across 24 diagnostic labels Provisional dermatologist vs final diagnosis, kappa 0.814 Almost perfect agreement across 24 diagnostic labels AUC for confidence score 0.866 Correct vs incorrect top-1 prediction Provisional vs final diagnosis concordance 82.9% (58/70) chi-square=30.23, p<0.0001 AI, artificial intelligence; AUC, area under the receiver-operating-characteristic curve; CI, confidence interval. Kappa values were calculated on the 24-category diagnostic label set. Performance by final diagnostic category In all the diagnoses, the macro-average sensitivity was 80.7%, and the macro-average specificity was 99.0%. Furthermore, the model was able to detect dermatophytoses with an accuracy of 91.7% and specificity of 93.1%. The sensitivity of the developed algorithm was 77.8% for acne vulgaris and 83.3% for onychomycosis; but atopic dermatitis or eczema had the lowest figure, i.e., 37.5%. The results for rare conditions cannot be justified due to the small number of cases diagnosed: only three or less were confirmed under each diagnosis code; therefore, they are used only in order to present the range of cases diagnosed (Table 4). Table 4. One-versus-rest performance of AI top-1 diagnosis by final confirmed category Final confirmed diagnosis n Sensitivity Specificity PPV NPV Macro-average, all categories - 80.7% 99.0% 69.2% 99.0% Dermatophytosis (tinea corporis) 12 91.7% 93.1% 73.3% 98.2% Acne vulgaris 9 77.8% 100.0% 100.0% 96.8% Atopic dermatitis/eczema 8 37.5% 100.0% 100.0% 92.5% Onychomycosis 6 83.3% 100.0% 100.0% 98.5% Pityriasis versicolor 5 60.0% 98.5% 75.0% 97.0% Fixed drug eruption 3 100.0% 100.0% 100.0% 100.0% Herpes zoster 3 66.7% 98.5% 66.7% 98.5% Melasma 3 100.0% 100.0% 100.0% 100.0% Psoriasis vulgaris 3 66.7% 98.5% 66.7% 98.5% Scabies 3 100.0% 100.0% 100.0% 100.0% Vitiligo 3 100.0% 97.0% 60.0% 100.0% Chronic urticaria 2 100.0% 100.0% 100.0% 100.0% Pemphigus vulgaris 2 50.0% 100.0% 100.0% 98.6% Pityriasis rosea 2 100.0% 100.0% 100.0% 100.0% Viral wart 2 100.0% 100.0% 100.0% 100.0% Other categories with <=1 confirmed case 1 each 0%-100.0% 98.6%-100.0% Not stable 98.6%-100.0% PPV, positive predictive value; NPV, negative predictive value. Categories with no confirmed cases were excluded from sensitivity calculations. Condition-specific values based on <=3 cases are imprecise and should not be used for comparative claims. Subgroup analysis In terms of top-1 accuracy, there was little difference between the good and adequate image-quality groups, 79.3% vs 75%. However, multiple lesions had slightly better accuracy than the single lesions, at 86.1% vs 70.6%, though no statistical comparisons were provided for groups. The lowest accuracy was recorded in the pediatric subgroup at 62.5% (n=8). Most marked differences were noted for the model confidence level: 12.5% accuracy in the low confidence, 79.4% accuracy in the moderate confidence, and 96.4% accuracy in the high confidence category (Table 5). Table 5. Subgroup analysis of AI top-1 diagnostic accuracy Domain Subgroup Correct/total Accuracy 95% CI Age group Paediatric (<18 years) 5/8 62.5% 30.6%-86.3% Adult 18-39 years 21/26 80.8% 62.1%-91.5% Adult 40-59 years 22/27 81.5% 63.3%-91.8% Older adult (>=60 years) 7/9 77.8% 45.3%-93.7% Sex Female 33/44 75.0% 60.6%-85.4% Male 22/26 84.6% 66.5%-93.9% Image quality Good 46/58 79.3% 67.2%-87.7% Adequate 9/12 75.0% 46.8%-91.1% Lesion count Multiple 31/36 86.1% 71.3%-93.9% Single 24/34 70.6% 53.8%-83.2% AI confidence Low (<0.60) 1/8 12.5% 2.2%-47.1% Moderate (0.60-0.79) 27/34 79.4% 63.2%-89.7% High (>=0.80) 27/28 96.4% 82.3%-99.4% AI, artificial intelligence; CI, confidence interval. No formal between-subgroup hypothesis tests were provided; values are descriptive.

This external validation of a locked multimodal dermatology model provides a measured estimate of performance in routine outpatient practice. The model achieved 78.6% top-1 accuracy and 92.9% top-3 accuracy, with substantial agreement between the first-ranked output and the dermatologist-confirmed final diagnosis. The central implication is not that the model can diagnose independently. Rather, the findings indicate that, when standardised photographs are interpreted alongside structured clinical information, the system can generate a clinically useful short differential diagnosis that merits clinician review. The most informative comparison is with the setting of previous studies, rather than with a single headline accuracy value. Recent reviews have largely evaluated AI for skin cancer detection using dermoscopic or selected clinical images, and they consistently describe major heterogeneity in disease spectrum, reference standards, image quality and outcome definitions [3-5,16,17]. Salinas et al. reported pooled AI sensitivity of 86.3% and specificity of 78.4% for skin cancer classification, with performance broadly comparable with expert dermatologists [5]. These estimates cannot be directly compared with the present top-1 accuracy or macro-average category metrics. Our model was tested on multiple outpatient dermatoses, including infectious and inflammatory disorders, for which the clinical distinction is often more nuanced than a binary malignant-versus-benign task. The 78.6% top-1 accuracy should therefore be interpreted as performance in a broader diagnostic environment, not as evidence of superiority or inferiority to cancer-focused systems. The need for this type of external validation is illustrated by Combalia et al. [6]. In the International Skin Imaging Collaboration challenge, the best algorithm achieved 82.0% balanced accuracy on HAM10000 but 58.8% on the more clinically representative BCN20000 dataset [6]. The outcome metric differs from the present study and no direct numerical comparison is appropriate. Nevertheless, the fall in performance after testing on a more representative dataset demonstrates how strongly image source, disease prevalence and case complexity can influence algorithmic performance. Our results add to this evidence by showing the performance of a locked model after transfer to a distinct outpatient population rather than after testing on a curated benchmark alone. The top-3 result deserves separate consideration. The correct diagnosis was included in the model’s first three suggestions in 92.9% of cases, suggesting that the system often retained the correct disorder even when the leading prediction was wrong. This is potentially useful for conditions with overlapping morphology, such as eczema, contact dermatitis, psoriasis and tinea, where clinicians work from a differential diagnosis rather than from an isolated image label. Barbieri et al. reported 90% accuracy and an AUC of 0.9646 for a multimodal model for leprosy diagnosis [11]. Although that top-1 accuracy exceeds the present estimate, their model addressed a single disease and cannot be considered a direct comparator for a multi-class outpatient cohort. Together with recent multimodal studies, including SkinGPT-4 and PanDerm, the present findings support the rationale for integrating clinical context with image information [9,11,13,14]. The condition-level results should be viewed as hypothesis-generating. Performance was most encouraging for dermatophytosis, acne and onychomycosis, whereas sensitivity was lower for atopic dermatitis or eczema. This pattern is clinically plausible. Dermatophytosis and onychomycosis may present with relatively recognisable morphology that is strengthened by information about symptom duration, distribution and recurrence. In contrast, eczema is a heterogeneous clinical construct with substantial overlap with contact dermatitis, psoriasis and tinea. A standard photograph may not capture scale, texture, lesion evolution, exposures or the full body distribution required to resolve these differentials. Similar limitations of image-dominant systems have been described in reviews of dermatology AI [9,10]. However, the small number of cases in each category means that no disease-specific performance claim should be made from this study. Confidence was associated with correctness and may have practical value, but it must not be mistaken for calibration. Concordant predictions had a mean confidence score 0.18 points higher than discordant predictions, and top-1 accuracy increased across confidence groups from 12.5% to 96.4%. The AUC of 0.866 indicates good discrimination in this cohort. A clinically sensible explanation is that the model is more certain when the photograph and history match a pattern represented in its development data. Yet models can remain highly confident in an incorrect output when presented with an uncommon disorder, an artefact, incomplete clinical context or a distribution shift. Reporting guidance therefore requires transparent description of model development, external validation, calibration and failure modes before confidence thresholds are incorporated into care pathways [15,18]. The subgroup findings reinforce the need for caution around equity and transportability. Accuracy was similar for good and adequate photographs, which may reflect the use of a standardised image-acquisition process. However, the lower paediatric estimate and the wide confidence intervals across several subgroups show that stable subgroup conclusions cannot be drawn from this sample. Daneshjou et al. demonstrated performance disparities across demographic and skin-tone groups in a diverse clinical image set, while Guo et al. documented underreporting and underrepresentation of diverse skin types in dermatology machine-learning research [7,8]. Real-world clinical image workflows also vary substantially from benchmark datasets, including in image acquisition and lesion selection [19]. This Indian cohort is valuable because it includes local presentations of fungal and inflammatory disorders, but it does not establish fairness across skin tone, disease severity, device type or socioeconomic background because these variables were not captured. The appropriate next step is evaluation of the model as an assistive tool. PanDerm improved clinician melanoma-diagnosis accuracy from 0.69 to 0.83 in a reader study [14], while Mijares et al. reported 47.1% top-diagnosis accuracy and 92.6% any-diagnosis accuracy for a dermatology-focused multimodal model evaluated in 68 biopsy-proven pigmented lesions [20]. Neither study is a direct comparator: both addressed different populations, endpoints and use cases. They do, however, reinforce an important point from the present study. A high-ranked differential may be more useful when it helps a clinician reconsider, prioritise or investigate a diagnosis than when it is treated as an autonomous answer. In practice, a discordant or low-confidence output should trigger review of history and examination findings, followed where appropriate by dermoscopy, microscopy, laboratory testing, biopsy, referral or follow-up. Prospective studies should test whether this workflow improves clinician performance, diagnostic delay, treatment decisions, referral appropriateness and patient outcomes [4,5]. The study is strengthened by its evaluation of a fixed multimodal model in a heterogeneous outpatient cohort, with standardised photographs, structured clinical information, ranked accuracy, agreement and confidence-based analyses. These features provide a more clinically meaningful assessment than a single benchmark accuracy estimate. Its limitations are equally important. The sample was small and single-centre, with sparse numbers in several diagnostic categories; category-specific estimates are therefore imprecise. The reference standard was the dermatologist-confirmed final diagnosis, but the source records did not identify histopathological, microbiological or laboratory confirmation for every case. In addition, the absence of detailed information on the model-development population, training labels, model version, calibration and handling of nondiagnostic images limits reproducibility and risk-of-bias assessment. Larger prospective multicentre studies should use transparent reporting, prespecified analyses, representative patient subgroups and composite reference standards where clinically appropriate.

A pre-developed multimodal AI model achieved 78.6% top-1 and 92.9% top-3 diagnostic accuracy in a heterogeneous outpatient dermatology cohort, with substantial agreement against final dermatologist-confirmed diagnoses. Prediction confidence was strongly associated with correctness and may help define a safety-oriented review pathway. The evidence supports clinical decision support, not independent diagnosis. Before use beyond research, the model requires transparent technical reporting, calibration assessment, fairness evaluation and prospective multicentre studies that measure patient-centred outcomes and failure modes

1. Yakupu A, Aimaier R, Yuan B, et al. The burden of skin and subcutaneous diseases: findings from the Global Burden of Disease Study 2019. Front Public Health. 2023;11:1145513. doi:10.3389/fpubh.2023.1145513. 2. Tian J, Zhang D, Yang Y, et al. Global epidemiology of atopic dermatitis: a comprehensive systematic analysis and modelling study. Br J Dermatol. 2023;190(1):55-61. doi:10.1093/bjd/ljad339. 3. Jones OT, Matin RN, van der Schaar M, et al. Artificial intelligence and machine learning algorithms for early detection of skin cancer in community and primary care settings: a systematic review. Lancet Digit Health. 2022;4(6):e466-e476. doi:10.1016/S2589-7500(22)00023-1. 4. Escale-Besa A, Vidal-Alaball J, Miro Catalina Q, et al. The use of artificial intelligence for skin disease diagnosis in primary care settings: a systematic review. Healthcare (Basel). 2024;12(12):1192. doi:10.3390/healthcare12121192. 5. Salinas MP, Sepulveda J, Hidalgo L, et al. A systematic review and meta-analysis of artificial intelligence versus clinicians for skin cancer diagnosis. NPJ Digit Med. 2024;7:125. doi:10.1038/s41746-024-01103-x. 6. Combalia M, Codella N, Rotemberg V, et al. Validation of artificial intelligence prediction models for skin cancer diagnosis using dermoscopy images: the 2019 International Skin Imaging Collaboration Grand Challenge. Lancet Digit Health. 2022;4(5):e330-e339. doi:10.1016/S2589-7500(22)00021-8. 7. Daneshjou R, Vodrahalli K, Novoa RA, et al. Disparities in dermatology AI performance on a diverse, curated clinical image set. Sci Adv. 2022;8(32):eabq6147. doi:10.1126/sciadv.abq6147. 8. Guo LN, Lee MS, Kassamali B, Mita C, Nambudiri VE. Bias in, bias out: underreporting and underrepresentation of diverse skin types in machine learning research for skin cancer detection: a scoping review. J Am Acad Dermatol. 2022;87(1):157-159. doi:10.1016/j.jaad.2021.06.884. 9. Luo N, Zhong X, Su L, Cheng Z, Ma W, Hao P. Artificial intelligence-assisted dermatology diagnosis: From unimodal to multimodal. Comput Biol Med. 2023;165:107413. doi:10.1016/j.compbiomed.2023.107413. 10. Patel RH, Foltz EA, Witkowski A, Ludzik J. Analysis of artificial intelligence-based approaches applied to non-invasive imaging for early detection of melanoma: a systematic review. Cancers (Basel). 2023;15(19):4694. doi:10.3390/cancers15194694. 11. Barbieri RR, Xu Y, Setian L, et al. Reimagining leprosy elimination with AI analysis of a combination of skin lesion images with demographic and clinical data. Lancet Reg Health Am. 2022;9:100192. doi:10.1016/j.lana.2022.100192. 12. Fernandes JRN, et al. Artificial intelligence on diagnostic aid of leprosy: a systematic literature review. J Clin Med. 2023;13(1):180. doi:10.3390/jcm13010180. 13. Zhou J, He X, Sun L, et al. Pre-trained multimodal large language model enhances dermatological diagnosis using SkinGPT-4. Nat Commun. 2024;15:5649. doi:10.1038/s41467-024-50043-3. 14. Yan S, Yu Z, Primiero C, et al. A multimodal vision foundation model for clinical dermatology. Nat Med. 2025;31(8):2691-2702. doi:10.1038/s41591-025-03747-y. 15. Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 2024;385:q902. doi:10.1136/bmj.q902. 16. Lin TL, Lee KH, Karmakar R, et al. Artificial intelligence-assisted dermatologic screening: epidemiology and clinical features of basal cell carcinoma, squamous cell carcinoma, seborrheic keratosis and actinic keratosis. Bioengineering (Basel). 2025;12(11):1258. doi:10.3390/bioengineering12111258. 17. Jairath N, Pahalyants V, Shah R, Weed J, Carucci JA, Criscito MC. Artificial intelligence in dermatology: a systematic review of its applications in melanoma and keratinocyte carcinoma diagnosis. Dermatol Surg. 2024;50(9):791-798. doi:10.1097/DSS.0000000000004223. 18. Moassefi M, Singh Y, Conte GM, et al. Checklist for reproducibility of deep learning in medical imaging. J Imaging Inform Med. 2024;37(4):1664-1673. doi:10.1007/s10278-024-01065-2. 19. Useini V, Tanadini-Lang S, Lohmeyer Q, et al. Automatized self-supervised learning for skin lesion screening. Sci Rep. 2024;14:12697. doi:10.1038/s41598-024-61681-4. 20. Mijares J, Jairath N, Zhang A, Que SKT. Validation of a dermatology-focused multimodal large language model in classification of pigmented skin lesions. Diagnostics (Basel). 2025;15(21):2808. doi:10.3390/diagnostics15212808.