Articles

Behavior Imaging, a Boise, Idaho company, uses a system called the Naturalistic Observation Diagnostic Assessment. In the privacy and comfort of their own homes, families use a smart phone app to capture and upload videos of their child’s behaviors in specified situations.

Clinicians watch the videos to make remote diagnoses. More recently, the company has started training AI-like algorithms to observe and categorize behaviors. Although, the algorithms would not diagnose the children, they might be used to point clinicians to specific behaviors that might otherwise have been missed.

Another use of AI-aided diagnosis is an autism screening tool created by Cognoa in Palo Alto California. The tool uses clinically validated artificial intelligence (AI) technology to aid physicians in diagnosing ASD in children between the ages of 18 and 72 months who are at risk of developmental delay.

AI thus reduces the quantum of work by a clinician and therefore speeds up the ASD diagnoses pipeline. We can conclude that AI is of great value in ASD research and the challenging problems related to ASD are ripe for the application of AI/ML technologies. EinNext R&D aims to develop promising screening tools to help decrease the length of time and cost required for diagnosis of ASD.

EinNext Biosciences

Autism Spectrum Disorders (ASD) comprise a group of neurodevelopmental abnormalities that begin in early childhood characterized by difficulty with social communication and interaction, restricted interests, and repetitive behaviors. According to the latest international data, the incidence of ASD has increased from 150:1 in 2000 to 36:1 in 2017, making it the leading cause of disability in children. With estimates of 2% of the children diagnosed with ASD in the US, researchers should investigate treatment methods, intensity, duration, and outcomes.

The Challenge:

Although the exact cause of ASD remains unclear, current studies suggests that it may be associated with genetic factors, abnormal brain structure and environmental factors.  Since autism spectrum disorder varies widely in symptoms and severity, with no specific medical test to diagnose the disorder, diagnosing ASD can be difficult. Healthcare providers diagnose the condition based on standardized assessments of the patient’s history and behavior. Although ASD can be diagnosed as early as 15- 18 months of age, many children do not receive a final diagnosis until they are in their adolescence or adulthood.

Early diagnosis and intervention are imperative for timely effective treatments to minimize symptoms, and to improve long-term outcomes related to cognition, language, adaptive behavior, daily living skills, and social behavior for children with ASD. Additionally, accurate identification is challenging as ASD is often enmeshed with other neurodevelopmental disorders, and medical comorbidities.

AI in Autism Research:

Since existing solutions for diagnosis of ASD are both resource-intensive and cost-intensive, Artificial intelligence (AI) is a potential solution to this issue. Autism spectrum disorder research has yet to leverage big data on the same scale as other fields. Advancements in easy, affordable data collection and analysis may soon make this a reality.

The high prevalence rate and heterogeneous nature of ASD have led some researchers to turn to machine learning over traditional statistical methods for data analysis. In the last decade, study of AI in Autism has shown a remarkable increase in trend. The availability of various machine learning toolkits, such as Hadoop, TensorFlow, Spark, and R, has led to unique opportunities for researchers to leverage machine learning algorithms.

 Hotspots and Research Fronts:

To improve screening and diagnosis, Machine learning algorithms—such as SVMs— have been used in ASD research. These models have been shown to improve the accuracy of diagnoses and provide insight into how different characteristics (such as standardized assessments, eye movement data, upper limb and general kinesthetic data, and neuroimaging data) can aid in the differential diagnosis of ASD. Standardized assessments that are currently in use have a potential for misdiagnoses, particularly when distinguishing one disorder from another. Machine learning procedures can discriminate individuals with ASD from individuals with attention deficit hyperactivity disorder (ADHD) with a high level of accuracy.

Machine learning has also been implemented in examining neuroimaging data. The symptom severity of individuals with ASD based on cortical thickness using support vector regression (SVR) and ENet penalized linear regression has been studied. Machine learning models, including RF, to analyze neuroimaging data for diagnostic classification purposes have also been implemented.

Longitudinal data captured at multiple points in development (8 and 14 months of age) for high-risk siblings has been used to increase accuracy of predicting ASD diagnosis at 36 months. Deep learning has been used to study predictors of challenging behavior and analyze neuroimaging in individuals with ASD.

While data from the electronic health records of decedents with an ASD were used to build a random forest classifier to examine the life-time health problems of those with ASD, the impact of parental age on the risk of developing ASD has been evaluated using logistic regression.

Machine learning has also been applied in ASD genetics research, allowing researchers to determine which genes are related to ASD. These findings highlight the potential of machine learning to improve understanding of the role genes play in the development of ASD.

One of the biggest challenges is the rarity and unpredictability of anomalies and failures, resulting in insufficient data to build a complete and accurate failure prediction model in real-world scenarios. In addition, the need for multiple models to address the different types of devices in a single plant can make operations and maintenance challenging. This is because these multiple models operate under different assumptions about the underlying patterns in each device’s data set.

To address these challenges, our team has developed an RL-based solution with the following features:

• The RL-based solution is designed to learn more effective anomaly detection strategies with a feasible amount of data.
• The solution employs a single anomaly detection model for monitoring the health of the entire factory, reducing the complexity of operations and maintenance.

The RLAD agent uses plant state as observations to predict anomaly scores for early detection of potential failures. The reward system is designed to avoid false detections and the agent’s predictions are used to indicate the plant’s health and root cause of abnormality. This helps operators schedule maintenance and minimize impact on production efficiency. The RLAD can also be applied to quality control and process monitoring in manufacturing.

Thus, reinforcement learning is a promising approach for solving manufacturing problems. Its ability to handle complex, dynamic systems and make decisions based on uncertain and changing information makes it well-suited for applications in production scheduling, predictive maintenance, and inventory control.

Software Team

Reinforcement learning is a subfield of artificial intelligence that focuses on enabling an agent to make decisions in an environment by continuously taking actions, observing the consequences, and learning from the rewards it receives for its actions. The ultimate goal of the agent is to learn a policy that maximizes its cumulative reward over time. In recent years, there has been growing interest in using reinforcement learning to address manufacturing problems.

We have recently applied RL in anomaly detection. Anomaly detection in manufacturing involves identifying any unusual or unexpected behavior in a production system, which can indicate potential problems such as equipment malfunctions, process disruptions, or quality issues. The goal of anomaly detection is to identify these deviations as early as possible so that corrective actions can be taken to minimize the impact on production and prevent costly downtime.

In RL-based anomaly detection, the agent acts as a detector and learns to identify anomalies in the factory process data through trial-and-error. The agent’s performance is evaluated based on its ability to accurately identify anomalies and receive rewards for successful detections. The agent’s policy for detecting anomalies is constantly updated and improved as it interacts with the data, making it capable of detecting anomalies in different types of devices within the factory. Thus, the agent serves as a single anomaly detector for the entire factory.

The reward equation in RL-based anomaly detection incentivizes the agent to predict anomalies as early as possible while minimizing false positive detections. The health of the factory process is indicated by the range of predicted anomaly scores, allowing the operator to quickly identify any deviations from the normal process. The root cause of these deviations is also displayed, providing the operator with crucial information to take the necessary corrective actions.