The Transformative Power of Artificial Neural Networks in Scientific Statistical Analysis
Hesam Mirmohammadi, Jafar Kolahi, Abbasali Khademi · Dental Hypotheses · 2024
In the rapidly growing landscape of academic research, the application of innovative and novel technologies such as artificial neural networks (ANNs) is essential, which have revolutionized the way researchers approach statistical analyses and data processing. At September 12, 2024 PubMed searched with the query "artificial neural network*"[All Fields] AND "statistical analysis"[All Fields]. Polynomial trend-line analysis of the results showed the number of articles was growing fast (y = 0.0681x2 − 272.53x + 272640, R2 = 0.8599). Results of author keywords co-accordance network analysis using VOSviewer 1.6.16 (www.vosviewer.com/, Leiden University Centre for Science and Technology Studies) showed in Figure 1.Figure 1: Results of author keywords co-accordance network analysis.However, an ANN can be made of a connected network of units or nodes called artificial neurons. Conceived as a physiology of biological neurons, an artificial neuron is a mathematical function that receives one or more inputs, applies weights to these inputs, sums them, and passes the sum through a non-linear function known as an activation function to produce an output.[1] ANNs are mainly beneficial when the associations between variables are not well understood or cannot be simply modeled with traditional linear statistical analysis. Some key advantages of ANNs for statistical analysis are: (1) Ability to model highly complex, nonlinear relationships without making assumption about the underlying data distribution, (2) Can handle large amount of data and make precise forecasts, (3) Can be used for classification, regression, clustering, dimensionally reduction, and other statistical tasks, and (4) Recurrent neural networks and long- and short-term memories can be used effectively for modeling temporal dependencies in time series data. Yet, ANNs have some inherent limitations that should be carefully considered. Some weaknesses of ANNs for statistical analyses are: (1) ANNs generally referred to “Black Box” because relationships between hidden layers of artificial neurons and decision making procedure are difficult to understand, (2) In contrast with traditional statistical methods, ANNs need a large amount of high-quality data for network training, and they are sensitive to noise, outliers, and missing values in the data, (3) Unlike traditional statistical models, ANNs do not provide statistical guarantees such as confidence intervals or hypothesis testing, which can be important for scientific inference and decision-making, (4) ANNs require intensive computational power, especially for large and complex models. However, to our knowledge, a combination of ANNs and traditional statistical methods can often provide the best approach for gaining insights from complex data. A prime example of the power of ANNs in statistical analysis can be found in genetics and genomics. ANNs had been employed successfully to analyze vast genomic datasets, uncovering previously unknown relationships between genetic sequences and their corresponding functions, which can open up new perspective in disease diagnosis, drug development, and personalized medicine, where the ability to make accurate predictions based on genetic information is of paramount importance.[2] In the same way, in the field of astrophysics, ANNs have been used successfully to analyze the big data generated by astronomical observatories and telescopes. Such ANN-derived information has the potential to revolutionize our fundamental knowledge of cosmology and the evolution of the cosmos.[3] Nevertheless, statistical theory behind the success of ANNs is still in the growing area of research. Yet, ANNs are powerful and flexible class of statistical models. As the theory and practice of ANNs continue to develop we can expect to use even more practical applications of ANNs in scientific statistical analysis.