Chemie  |  Biochemie  |  Medizin

 

Himani Patel, 2005 | Küttigen, AG
Katie Richard, 2005 | Suhr, AG

 

Autism is a complex condition that presents itself in many ways, with symptoms as varied as the individual affected by it. Despite this, the way it is currently diagnosed and thought about heavily relies on male symptomology. To correct this inequality and improve diagnostics for women, we conducted a literature review of 30 papers regarding sex/gender differences in the symptomology of autism as well as created a Machine Learning model to screen for autism in females, to improve our knowledge of the issue and apply it to improving autism screening for women.

Introduction

1. What does current research reveal about the clinical and behavioral characteristics of Autism Spectrum Condition (ASC) in females
2. Can a gender-tailored machine learning model improve preliminary diagnostic accuracy for autism in women and help mitigate gender bias in clinical assessments?

Methods

Literature Review
We searched for papers using the terms «autism», «differences», and «sex» or «gender,» selecting 15 papers with numerical data, excluding meta-analyses and biological studies. The 30 papers were analyzed and results were recorded in categories: Social Competence, Comorbidities, Restrictive Repetitive Behaviors, Language Ability, Externalizing Behavior, Camouflaging, and Cognitive Ability. Additional details like publishing year, title, sample size, and male-to-female ratio were also noted for reliability.
Model
We found one suitable publicly accessible dataset. After studying a Python program based on this data, we reduced the sample size by excluding male participants, leaving 300 female instances. The model structure focused on selecting the best three algorithms from 11, evaluated using ROC AUC and F1-Score. The chosen algorithms—Decision Tree Classifier, Gaussian Naïve Bayes, and Gradient Boosting Classifier—were weighted to emphasize questions more relevant to women.

Results

The model evaluation revealed an ROC AUC Score consistently between 0.87-0.89 and an F1-Score around 0.59-0.63. The manual enhancement of the AQ-10 questions led to further gains in both ROC and F1 scores. As a comparison, the existing AQ-10 screening tests demonstrate ROC AUC scores of around 0.52-0.58, depending on the data composition, so it outperforms the by hand review AQ-10. Correlation analysis, especially Correlation Matrices, revealed moderate linear associations with the target variable, with stronger correlations observed in males compared to females.

The results of the Literature Review imply that the typical profile of an autistic female would be as following:
– Have an increased rate of camouflaging
– Have less overt conflicts than males
– Be more likely to internalize issues
– Have relationships similar to neurotypical females, but have more conflict
– Have less apparent/less stereotypical RRBs
– Be better at social reciprocation

Discussion

The model outperforms the AQ-10 screening tool, with an ROC AUC Score above 0.8 compared to around 0.55 for AQ-10. It shows high potential as an automated screening tool for subgroups, such as women. With an ROC Score above 0.8, the model demonstrated good class differentiation, while an F1-Score of 0.6 indicated a balanced precision-recall trade-off, despite some errors. Imbalanced datasets likely biased the model toward the majority class. Key challenges included insufficient data and lack of domain knowledge for scaling, highlighting the need for prior research. Integrating domain knowledge, possibly through genetic algorithms, and algorithm-specific analyses could enhance performance. Larger datasets would also improve evaluation and the review process.
The review process improved our understanding of current assumptions regarding the presentation of autism between genders by comparing studies based on key factors. Despite this, reliance on Google Scholar may have misrepresented the relevance of certain studies and thus skewed the results.
Additionally, the scope lacked a detailed analysis of age-related behavioral differences across autistic populations, which may have affected the data.

Conclusions

Sex/gender differences in autistic traits have no consensus on key distinctions. According to the Literature Review, an optimal autism questionnaire for women and girls should focus on camouflaging behaviors, internalizing issues, relational conflicts, and focus on repetitive behaviors and traits more common in females. Clinical diagnoses should incorporate a longer term observation for higher accuracy. Machine learning models trained on subgroup data could improve screening tools by incorporating domain knowledge and adjusting question weightings.

 

 

Würdigung durch die Expertin

Lea Burch

Katie Richard und Himani Patel widmen sich in ihrer Arbeit der diagnostisch schwierigen und wissenschaftlich unterbeleuchteten Problematik der Erkennung der Autismus-Spektrum-Störung bei Frauen. Mit grossem Engagement analysieren sie das Thema zunächst in einer fundierten Literaturrecherche und untersuchen anschliessend, ob mithilfe eines Machine-Learning-Ansatzes ein gängiges Screening-Instrument gezielter auf weibliche Merkmale zugeschnitten werden kann. Die Arbeit überzeugt durch klare Struktur, gute Verständlichkeit und einen gelungenen Brückenschlag zwischen Theorie und Technik.

Prädikat:

sehr gut

 

 

 

Alte Kantonsschule Aarau
Lehrerin: Martina Vàzquez