RESUMO
Large language models (LLMs) are generating interest in medical settings. For example, LLMs can respond coherently to medical queries by providing plausible differential diagnoses based on clinical notes. However, there are many questions to explore, such as evaluating differences between open- and closed-source LLMs as well as LLM performance on queries from both medical and non-medical users. In this study, we assessed multiple LLMs, including Llama-2-chat, Vicuna, Medllama2, Bard/Gemini, Claude, ChatGPT3.5, and ChatGPT-4, as well as non-LLM approaches (Google search and Phenomizer) regarding their ability to identify genetic conditions from textbook-like clinician questions and their corresponding layperson translations related to 63 genetic conditions. For open-source LLMs, larger models were more accurate than smaller LLMs: 7b, 13b, and larger than 33b parameter models obtained accuracy ranges from 21%-49%, 41%-51%, and 54%-68%, respectively. Closed-source LLMs outperformed open-source LLMs, with ChatGPT-4 performing best (89%-90%). Three of 11 LLMs and Google search had significant performance gaps between clinician and layperson prompts. We also evaluated how in-context prompting and keyword removal affected open-source LLM performance. Models were provided with 2 types of in-context prompts: list-type prompts, which improved LLM performance, and definition-type prompts, which did not. We further analyzed removal of rare terms from descriptions, which decreased accuracy for 5 of 7 evaluated LLMs. Finally, we observed much lower performance with real individuals' descriptions; LLMs answered these questions with a maximum 21% accuracy.
RESUMO
Artificial intelligence (AI) for facial diagnostics is increasingly used in the genetics clinic to evaluate patients with potential genetic conditions. Current approaches focus on one type of AI called Deep Learning (DL). While DL- based facial diagnostic platforms have a high accuracy rate for many conditions, less is understood about how this technology assesses and classifies (categorizes) images, and how this compares to humans. To compare human and computer attention, we performed eye-tracking analyses of geneticist clinicians (n = 22) and non-clinicians (n = 22) who viewed images of people with 10 different genetic conditions, as well as images of unaffected individuals. We calculated the Intersection-over-Union (IoU) and Kullback-Leibler divergence (KL) to compare the visual attentions of the two participant groups, and then the clinician group against the saliency maps of our deep learning classifier. We found that human visual attention differs greatly from DL model's saliency results. Averaging over all the test images, IoU and KL metric for the successful (accurate) clinician visual attentions versus the saliency maps were 0.15 and 11.15, respectively. Individuals also tend to have a specific pattern of image inspection, and clinicians demonstrate different visual attention patterns than non-clinicians (IoU and KL of clinicians versus non-clinicians were 0.47 and 2.73, respectively). This study shows that humans (at different levels of expertise) and a computer vision model examine images differently. Understanding these differences can improve the design and use of AI tools, and lead to more meaningful interactions between clinicians and AI technologies.
Assuntos
Inteligência Artificial , Computadores , Humanos , Simulação por ComputadorRESUMO
Artificial intelligence (AI) is increasingly used in genomics research and practice, and generative AI has garnered significant recent attention. In clinical applications of generative AI, aspects of the underlying datasets can impact results, and confounders should be studied and mitigated. One example involves the facial expressions of people with genetic conditions. Stereotypically, Williams (WS) and Angelman (AS) syndromes are associated with a "happy" demeanor, including a smiling expression. Clinical geneticists may be more likely to identify these conditions in images of smiling individuals. To study the impact of facial expression, we analyzed publicly available facial images of approximately 3500 individuals with genetic conditions. Using a deep learning (DL) image classifier, we found that WS and AS images with non-smiling expressions had significantly lower prediction probabilities for the correct syndrome labels than those with smiling expressions. This was not seen for 22q11.2 deletion and Noonan syndromes, which are not associated with a smiling expression. To further explore the effect of facial expressions, we computationally altered the facial expressions for these images. We trained HyperStyle, a GAN-inversion technique compatible with StyleGAN2, to determine the vector representations of our images. Then, following the concept of InterfaceGAN, we edited these vectors to recreate the original images in a phenotypically accurate way but with a different facial expression. Through online surveys and an eye-tracking experiment, we examined how altered facial expressions affect the performance of human experts. We overall found that facial expression is associated with diagnostic accuracy variably in different genetic conditions.
Assuntos
Expressão Facial , Humanos , Aprendizado Profundo , Inteligência Artificial , Genética Médica/métodos , Síndrome de Williams/genéticaRESUMO
Deep learning (DL) is applied in many biomedical areas. We performed a scoping review on DL in medical genetics. We first assessed 14,002 articles, of which 133 involved DL in medical genetics. DL in medical genetics increased rapidly during the studied period. In medical genetics, DL has largely been applied to small data sets of affected individuals (mean = 95, median = 29) with genetic conditions (71 different genetic conditions were studied; 24 articles studied multiple conditions). A variety of data types have been used in medical genetics, including radiologic (20%), ophthalmologic (14%), microscopy (8%), and text-based data (4%); the most common data type was patient facial photographs (46%). DL authors and research subjects overrepresent certain geographic areas (United States, Asia, and Europe). Convolutional neural networks (89%) were the most common method. Results were compared with human performance in 31% of studies. In total, 51% of articles provided data access; 16% released source code. To further explore DL in genomics, we conducted an additional analysis, the results of which highlight future opportunities for DL in medical genetics. Finally, we expect DL applications to increase in the future. To aid data curation, we evaluated a DL, random forest, and rule-based classifier at categorizing article abstracts.
Assuntos
Aprendizado Profundo , Genética Médica , Ásia , Genômica , Humanos , Redes Neurais de ComputaçãoRESUMO
Developing new software tools for analysis of large-scale biological data is a key component of advancing modern biomedical research. Scientific reproduction of published findings requires running computational tools on data generated by such studies, yet little attention is presently allocated to the installability and archival stability of computational software tools. Scientific journals require data and code sharing, but none currently require authors to guarantee the continuing functionality of newly published tools. We have estimated the archival stability of computational biology software tools by performing an empirical analysis of the internet presence for 36,702 omics software resources published from 2005 to 2017. We found that almost 28% of all resources are currently not accessible through uniform resource locators (URLs) published in the paper they first appeared in. Among the 98 software tools selected for our installability test, 51% were deemed "easy to install," and 28% of the tools failed to be installed at all because of problems in the implementation. Moreover, for papers introducing new software, we found that the number of citations significantly increased when authors provided an easy installation process. We propose for incorporation into journal policy several practical solutions for increasing the widespread installability and archival stability of published bioinformatics software.
Assuntos
Biologia Computacional/métodos , Disseminação de Informação/métodos , Armazenamento e Recuperação da Informação/métodos , Pesquisa Biomédica , Bases de Dados Factuais , Humanos , Internet , Software/tendênciasRESUMO
Caesarean delivery rates are increasing in many Asian countries. This study investigated the effects of caesarean section on breastfeeding practices from delivery to twelve months postpartum. A prospective cohort study was conducted on 2030 pregnant women recruited from three cities in Vietnam during 2015-2017. The overall caesarean rate was 38.1%. Mothers who underwent caesarean section were more likely to give prelacteal feeds to their infants (adjusted odds ratio (OR) 13.91, 95% confidence interval (CI) 10.52-18.39) and as a result have lower rates of early initiation of breastfeeding (adjusted OR 0.04, 95%CI 0.02-0.05). Having a caesarean section reduced the likelihood of (any, predominant and exclusive) breastfeeding from discharge to 6 months postpartum. After 1 year, the any breastfeeding rate was still lower in the caesarean delivery (70.2%) compared with the vaginal delivery group (72.9%), p = .232. Vietnamese women who give birth by caesarean section need extra support to initiate and maintain breastfeeding.IMPACT STATEMENTWhat is already known on this subject? Early initiation of breastfeeding, and 'exclusive' or 'predominant' breastfeeding rates at discharge are lower in mothers delivering by caesarean section compared to vaginal delivery. Prelacteal feeding rates are higher following caesarean section. However, the association between 'any' breastfeeding duration and caesarean delivery has not been established.What the results of this study add? This study showed that caesarean delivery reduced all breastfeeding rates from discharge to six months and any breastfeeding rate at 12 months postpartum in Vietnamese women.What the implications are of these findings for clinical practice and/or further research? Further breastfeeding interventions are needed during the postpartum period for mothers who deliver by caesarean section.
Assuntos
Aleitamento Materno/estatística & dados numéricos , Cesárea/efeitos adversos , Adulto , Estudos de Casos e Controles , Feminino , Humanos , Lactente , Recém-Nascido , Período Pós-Parto , Gravidez , Estudos Prospectivos , VietnãRESUMO
Meta-analysis strategies have become critical to augment power of genome-wide association studies (GWAS). To reduce genotyping or sequencing cost, many studies today utilize shared controls, and these individuals can inadvertently overlap among multiple studies. If these overlapping individuals are not taken into account in meta-analysis, they can induce spurious associations. In this article, we propose a general framework for adjusting association statistics to account for overlapping subjects within a meta-analysis. The key idea of our method is to transform the covariance structure of the data, so it can be used in downstream analyses. As a result, the strategy is very flexible and allows a wide range of meta-analysis methods, such as the random effects model, to account for overlapping subjects. Using simulations and real datasets, we demonstrate that our method has utility in meta-analyses of GWAS, as well as in a multi-tissue mouse expression quantitative trait loci (eQTL) study where our method increases the number of discovered eQTL by up to 19% compared with existing methods.
Assuntos
Doença/genética , Estudo de Associação Genômica Ampla/métodos , Metanálise como Assunto , Polimorfismo de Nucleotídeo Único/genética , Locos de Características Quantitativas/genética , Animais , Estudos de Casos e Controles , Perfilação da Expressão Gênica , Humanos , Camundongos , Modelos TeóricosRESUMO
MOTIVATION: There is recent interest in using gene expression data to contextualize findings from traditional genome-wide association studies (GWAS). Conditioned on a tissue, expression quantitative trait loci (eQTLs) are genetic variants associated with gene expression, and eGenes are genes whose expression levels are associated with genetic variants. eQTLs and eGenes provide great supporting evidence for GWAS hits and important insights into the regulatory pathways involved in many diseases. When a significant variant or a candidate gene identified by GWAS is also an eQTL or eGene, there is strong evidence to further study this variant or gene. Multi-tissue gene expression datasets like the Gene Tissue Expression (GTEx) data are used to find eQTLs and eGenes. Unfortunately, these datasets often have small sample sizes in some tissues. For this reason, there have been many meta-analysis methods designed to combine gene expression data across many tissues to increase power for finding eQTLs and eGenes. However, these existing techniques are not scalable to datasets containing many tissues, like the GTEx data. Furthermore, these methods ignore a biological insight that the same variant may be associated with the same gene across similar tissues. RESULTS: We introduce a meta-analysis model that addresses these problems in existing methods. We focus on the problem of finding eGenes in gene expression data from many tissues, and show that our model is better than other types of meta-analyses. AVAILABILITY AND IMPLEMENTATION: Source code is at https://github.com/datduong/RECOV . CONTACT: eeskin@cs.ucla.edu or datdb@cs.ucla.edu. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.
Assuntos
Biologia Computacional/métodos , Variação Genética , Locos de Características Quantitativas , Software , Perfilação da Expressão Gênica/métodos , Estudo de Associação Genômica Ampla/métodos , Humanos , Metanálise como Assunto , Modelos GenéticosRESUMO
MOTIVATION: Expression quantitative trait loci (eQTLs) are genetic variants that affect gene expression. In eQTL studies, one important task is to find eGenes or genes whose expressions are associated with at least one eQTL. The standard statistical method to determine whether a gene is an eGene requires association testing at all nearby variants and the permutation test to correct for multiple testing. The standard method however does not consider genomic annotation of the variants. In practice, variants near gene transcription start sites (TSSs) or certain histone modifications are likely to regulate gene expression. In this article, we introduce a novel eGene detection method that considers this empirical evidence and thereby increases the statistical power. RESULTS: We applied our method to the liver Genotype-Tissue Expression (GTEx) data using distance from TSSs, DNase hypersensitivity sites, and six histone modifications as the genomic annotations for the variants. Each of these annotations helped us detected more candidate eGenes. Distance from TSS appears to be the most important annotation; specifically, using this annotation, our method discovered 50% more candidate eGenes than the standard permutation method. CONTACT: buhm.han@amc.seoul.kr or eeskin@cs.ucla.edu.
Assuntos
Genômica , Variação Genética , Genótipo , Polimorfismo de Nucleotídeo Único , Locos de Características QuantitativasRESUMO
BACKGROUND: To determine the impact of cataract surgery on vision-related quality of life (VRQOL) and examine the association between objective visual measures and change in VRQOL after surgery among bilateral cataract patients in Ho Chi Minh City, Vietnam. METHODS: A cohort of older patients with bilateral cataract was assessed one week before and one to three months after first eye or both eye cataract surgery. Visual measures including visual acuity, contrast sensitivity and stereopsis were obtained. Vision-related quality of life was assessed using the NEI VFQ-25. Descriptive analyses and a generalized linear estimating equation (GEE) analysis were undertaken to measure change in VRQOL after surgery. RESULTS: Four hundred and thirteen patients were assessed before cataract surgery and 247 completed the follow-up assessment one to three months after first or both eye cataract surgery. Overall, VRQOL significantly improved after cataract surgery (p < 0.001) particularly after both eye surgeries. Binocular contrast sensitivity (p < 0.001) and stereopsis (p < 0.001) were also associated with change in VRQOL after cataract surgery. Visual acuity was not associated with VRQOL. CONCLUSIONS: Cataract surgery significantly improved VRQOL among bilateral cataract patients in Vietnam. Contrast sensitivity as well as stereopsis, rather than visual acuity significantly affected VRQOL after cataract surgery.
Assuntos
Extração de Catarata/psicologia , Qualidade de Vida , Idoso , Catarata/complicações , Catarata/psicologia , Extração de Catarata/estatística & dados numéricos , Sensibilidades de Contraste , Feminino , Humanos , Masculino , Pessoa de Meia-Idade , Estudos Prospectivos , Qualidade de Vida/psicologia , Inquéritos e Questionários , Vietnã/epidemiologia , Testes Visuais , Acuidade VisualRESUMO
BACKGROUND: Depression is common among older populations with cataract. However, the impact of cataract surgery on depression in both developed and developing countries remains unclear. The aim of this study is to determine the impact of cataract surgery on depressive symptoms and to examine the association between objective visual measures and change in depressive symptoms after surgery among a Vietnamese population in Ho Chi Minh City. METHODS: A cohort of older patients with bilateral cataract were assessed the week before and one to three months after first eye surgery only or first- and second-eye cataract surgeries. Visual measures including visual acuity, contrast sensitivity, and stereopsis were obtained. Depressive symptoms were assessed using the 20-item Center for Epidemiological Studies-Depression Scale (CES-D). Descriptive analyses and a generalized estimating equations (GEE) analysis were undertaken to determine the impact of cataract surgery on depressive symptoms. RESULTS: Four hundred and thirteen participants were recruited into the study before cataract surgery. Two hundred and forty-seven completed the follow-up assessment after surgery. There was a significant decrease (improvement) of one point in the depressive symptoms score (p = 0.04) after cataract surgery, after accounting for potential confounding factors. In addition, females reported a significantly greater decrease (improvement) of two points in depressive symptom scores (p = 0.01), compared to males. However, contrast sensitivity, visual acuity, and stereopsis were not significantly associated with change in depressive symptoms scores. First-eye cataract surgery or both-eye cataract surgery did not modify the change in depressive symptoms score. CONCLUSION: There was a small but significant improvement in depressive symptoms score after cataract surgery for an older population in Vietnam.
Assuntos
Extração de Catarata , Catarata , Depressão , Idoso , Catarata/diagnóstico , Catarata/psicologia , Extração de Catarata/métodos , Extração de Catarata/psicologia , Fatores de Confusão Epidemiológicos , Sensibilidades de Contraste , Depressão/diagnóstico , Depressão/fisiopatologia , Percepção de Profundidade , Feminino , Humanos , Vida Independente , Masculino , Pessoa de Meia-Idade , Avaliação de Resultados em Cuidados de Saúde , Período Pós-Operatório , Inquéritos e Questionários , Resultado do Tratamento , Acuidade VisualRESUMO
Large-language models like ChatGPT have recently received a great deal of attention. One area of interest pertains to how these models could be used in biomedical contexts, including related to human genetics. To assess one facet of this, we compared the performance of ChatGPT versus human respondents (13,642 human responses) in answering 85 multiple-choice questions about aspects of human genetics. Overall, ChatGPT did not perform significantly differently (p = 0.8327) than human respondents; ChatGPT was 68.2% accurate, compared to 66.6% accuracy for human respondents. Both ChatGPT and humans performed better on memorization-type questions versus critical thinking questions (p < 0.0001). When asked the same question multiple times, ChatGPT frequently provided different answers (16% of initial responses), including for both initially correct and incorrect answers, and gave plausible explanations for both correct and incorrect answers. ChatGPT's performance was impressive, but currently demonstrates significant shortcomings for clinical or other high-stakes use. Addressing these limitations will be important to guide adoption in real-life situations.
Assuntos
Inteligência Artificial , Genética Humana , HumanosRESUMO
Importance: The lack of standardized genetics training in pediatrics residencies, along with a shortage of medical geneticists, necessitates innovative educational approaches. Objective: To compare pediatric resident recognition of Kabuki syndrome (KS) and Noonan syndrome (NS) after 1 of 4 educational interventions, including generative artificial intelligence (AI) methods. Design, Setting, and Participants: This comparative effectiveness study used generative AI to create images of children with KS and NS. From October 1, 2022, to February 28, 2023, US pediatric residents were provided images through a web-based survey to assess whether these images helped them recognize genetic conditions. Interventions: Participants categorized 20 images after exposure to 1 of 4 educational interventions (text-only descriptions, real images, and 2 types of images created by generative AI). Main Outcomes and Measures: Associations between educational interventions with accuracy and self-reported confidence. Results: Of 2515 contacted pediatric residents, 106 and 102 completed the KS and NS surveys, respectively. For KS, the sensitivity of text description was 48.5% (128 of 264), which was not significantly different from random guessing (odds ratio [OR], 0.94; 95% CI, 0.69-1.29; P = .71). Sensitivity was thus compared for real images vs random guessing (60.3% [188 of 312]; OR, 1.52; 95% CI, 1.15-2.00; P = .003) and 2 types of generative AI images vs random guessing (57.0% [212 of 372]; OR, 1.32; 95% CI, 1.04-1.69; P = .02 and 59.6% [193 of 324]; OR, 1.47; 95% CI, 1.12-1.94; P = .006) (denominators differ according to survey responses). The sensitivity of the NS text-only description was 65.3% (196 of 300). Compared with text-only, the sensitivity of the real images was 74.3% (205 of 276; OR, 1.53; 95% CI, 1.08-2.18; P = .02), and the sensitivity of the 2 types of images created by generative AI was 68.0% (204 of 300; OR, 1.13; 95% CI, 0.77-1.66; P = .54) and 71.0% (247 of 328; OR, 1.30; 95% CI, 0.92-1.83; P = .14). For specificity, no intervention was statistically different from text only. After the interventions, the number of participants who reported being unsure about important diagnostic facial features decreased from 56 (52.8%) to 5 (7.6%) for KS (P < .001) and 25 (24.5%) to 4 (4.7%) for NS (P < .001). There was a significant association between confidence level and sensitivity for real and generated images. Conclusions and Relevance: In this study, real and generated images helped participants recognize KS and NS; real images appeared most helpful. Generated images were noninferior to real images and could serve an adjunctive role, particularly for rare conditions.
Assuntos
Anormalidades Múltiplas , Inteligência Artificial , Face/anormalidades , Doenças Hematológicas , Aprendizagem , Doenças Vestibulares , Humanos , Criança , Reconhecimento Psicológico , EscolaridadeRESUMO
UNLABELLED: Recent technological developments in measuring genetic variation have ushered in an era of genome-wide association studies which have discovered many genes involved in human disease. Current methods to perform association studies collect genetic information and compare the frequency of variants in individuals with and without the disease. Standard approaches do not take into account any information on whether or not a given variant is likely to have an effect on the disease. We propose a novel method for computing an association statistic which takes into account prior information. Our method improves both power and resolution by 8% and 27%, respectively, over traditional methods for performing association studies when applied to simulations using the HapMap data. Advantages of our method are that it is as simple to apply to association studies as standard methods, the results of the method are interpretable as the method reports p-values, and the method is optimal in its use of prior information in regards to statistical power. AVAILABILITY: The method presented herein is available at http://masa.cs.ucla.edu.
Assuntos
Biologia Computacional/métodos , Estudo de Associação Genômica Ampla , Frequência do Gene , Variação Genética , Projeto HapMap , Humanos , Funções Verossimilhança , Polimorfismo de Nucleotídeo ÚnicoRESUMO
Large-language models like ChatGPT have recently received a great deal of attention. To assess ChatGPT in the field of genetics, we compared its performance to human respondents in answering genetics questions (involving 13,636 responses) that had been posted on social media platforms starting in 2021. Overall, ChatGPT did not perform significantly differently than human respondents, but did significantly better on memorization-type questions versus critical thinking questions, frequently provided different answers when asked questions multiple times, and provided plausible explanations for both correct and incorrect answers.
RESUMO
Artificial intelligence (AI) is used in an increasing number of areas, with recent interest in generative AI, such as using ChatGPT to generate programming code or DALL-E to make illustrations. We describe the use of generative AI in medical education. Specifically, we sought to determine whether generative AI could help train pediatric residents to better recognize genetic conditions. From publicly available images of individuals with genetic conditions, we used generative AI methods to create new images, which were checked for accuracy with an external classifier. We selected two conditions for study, Kabuki (KS) and Noonan (NS) syndromes, which are clinically important conditions that pediatricians may encounter. In this study, pediatric residents completed 208 surveys, where they each classified 20 images following exposure to one of 4 possible educational interventions, including with and without generative AI methods. Overall, we find that generative images perform similarly but appear to be slightly less helpful than real images. Most participants reported that images were useful, although real images were felt to be more helpful. We conclude that generative AI images may serve as an adjunctive educational tool, particularly for less familiar conditions, such as KS.
RESUMO
Purpose: To describe the relationships between foveal structure and visual function in a cohort of individuals with foveal hypoplasia (FH) and to estimate FH grade and visual acuity using a deep learning classifier. Design: Retrospective cohort study and experimental study. Participants: A total of 201 patients with FH were evaluated at the National Eye Institute from 2004 to 2018. Methods: Structural components of foveal OCT scans and corresponding clinical data were analyzed to assess their contributions to visual acuity. To automate FH scoring and visual acuity correlations, we evaluated the following 3 inputs for training a neural network predictor: (1) OCT scans, (2) OCT scans and metadata, and (3) real OCT scans and fake OCT scans created from a generative adversarial network. Main Outcome Measures: The relationships between visual acuity outcomes and determinants, such as foveal morphology, nystagmus, and refractive error. Results: The mean subject age was 24.4 years (range, 1-73 years; standard deviation = 18.25 years) at the time of OCT imaging. The mean best-corrected visual acuity (n = 398 eyes) was equivalent to a logarithm of the minimal angle of resolution (LogMAR) value of 0.75 (Snellen 20/115). Spherical equivalent refractive error (SER) ranged from -20.25 diopters (D) to +13.63 D with a median of +0.50 D. The presence of nystagmus and a high-LogMAR value showed a statistically significant relationship (P < 0.0001). The participants whose SER values were farther from plano demonstrated higher LogMAR values (n = 382 eyes). The proportion of patients with nystagmus increased with a higher FH grade. Variability in SER with grade 4 (range, -20.25 D to +13.00 D) compared with grade 1 (range, -8.88 D to +8.50 D) was statistically significant (P < 0.0001). Our neural network predictors reliably estimated the FH grading and visual acuity (correlation to true value > 0.85 and > 0.70, respectively) for a test cohort of 37 individuals (98 OCT scans). Training the predictor on real OCT scans with metadata and fake OCT scans improved the accuracy over the model trained on real OCT scans alone. Conclusions: Nystagmus and foveal anatomy impact visual outcomes in patients with FH, and computational algorithms reliably estimate FH grading and visual acuity.
RESUMO
Deep learning (DL) and other types of artificial intelligence (AI) are increasingly used in many biomedical areas, including genetics. One frequent use in medical genetics involves evaluating images of people with potential genetic conditions to help with diagnosis. A central question involves better understanding how AI classifiers assess images compared to humans. To explore this, we performed eye-tracking analyses of geneticist clinicians and non-clinicians. We compared results to DL-based saliency maps. We found that human visual attention when assessing images differs greatly from the parts of images weighted by the DL model. Further, individuals tend to have a specific pattern of image inspection, and clinicians demonstrate different visual attention patterns than non-clinicians.
RESUMO
Neural networks have shown strong potential in research and in healthcare. Mainly due to the need for large datasets, these applications have focused on common medical conditions, where more data are typically available. Leveraging publicly available data, we trained a neural network classifier on images of rare genetic conditions with skin findings. We used approximately 100 images per condition to classify 6 different genetic conditions. We analyzed both preprocessed images that were cropped to show only the skin lesions as well as more complex images showing features such as the entire body segment, the person, and/or the background. The classifier construction process included attribution methods to visualize which pixels were most important for computer-based classification. Our classifier was significantly more accurate than pediatricians or medical geneticists for both types of images and suggests steps for further research involving clinical scenarios and other applications.
RESUMO
Background: In medical genetics, one application of neural networks is the diagnosis of genetic diseases based on images of patient faces. While these applications have been validated in the literature with primarily pediatric subjects, it is not known whether these applications can accurately diagnose patients across a lifespan. We aimed to extend previous works to determine whether age plays a factor in facial diagnosis as well as to explore other factors that may contribute to the overall diagnostic accuracy. Methods: To investigate this, we chose two relatively common conditions, Williams syndrome and 22q11.2 deletion syndrome. We built a neural network classifier trained on images of affected and unaffected individuals of different ages and compared classifier accuracy to clinical geneticists. We analyzed the results of saliency maps and the use of generative adversarial networks to boost accuracy. Results: Our classifier outperformed clinical geneticists at recognizing face images of these two conditions within each of the age groups (the performance varied between the age groups): 1) under 2 years old, 2) 2-9 years old, 3) 10-19 years old, 4) 20-34 years old, and 5) ≥35 years old. The overall accuracy improvement by our classifier over the clinical geneticists was 15.5 and 22.7% for Williams syndrome and 22q11.2 deletion syndrome, respectively. Additionally, comparison of saliency maps revealed that key facial features learned by the neural network differed with respect to age. Finally, joint training real images with multiple different types of fake images created by a generative adversarial network showed up to 3.25% accuracy gain in classification accuracy. Conclusion: The ability of clinical geneticists to diagnose these conditions is influenced by the age of the patient. Deep learning technologies such as our classifier can more accurately identify patients across the lifespan based on facial features. Saliency maps of computer vision reveal that the syndromic facial feature attributes change with the age of the patient. Modest improvements in the classifier accuracy were observed when joint training was carried out with both real and fake images. Our findings highlight the need for a greater focus on age as a confounder in facial diagnosis.