Developing Prognosis Tools to Identify Learning Difficulties in Children Using Machine Learning Technologies

Abstract

The Mental Attributes Profiling System was developed in 2002 (Laouris and Makris, Proceedings of multilingual & cross-cultural perspectives on Dyslexia, Omni Shoreham Hotel, Washington, D.C, 2002), to provide a multimodal evaluation of the learning potential and abilities of young children's brains. The method is based on the assessment of non-verbal abilities using video-like interfaces and was compared to more established methodologies in (Papadopoulos, Laouris, Makris, Proceedings of IDA 54th annual conference, San Diego, 2003), such as the Wechsler Intelligence Scale for Children (Watkins et al., Psychol Sch 34(4):309-319, 1997). To do so, various tests have been applied to a population of 134 children aged 7-12 years old. This paper addresses the issue of identifying a minimal set of variables that are able to accurately predict the learning abilities of a given child. The use of Machine Learning technologies to do this provides the advantage of making no prior assumptions about the nature of the data and eliminating natural bias associated with data processing carried out by humans. Kohonen's Self Organising Maps (Kohonen, Biol Cybern 43:59-69, 1982) algorithm is able to split a population into groups based on large and complex sets of observations. Once the population is split, the individual groups can then be probed for their defining characteristics providing insight into the rationale of the split. The characteristics identified form the basis of classification systems that are able to accurately predict which group an individual will belong to, using only a small subset of the tests available. The specifics of this methodology are detailed herein, and the resulting classification systems provide an effective tool to prognose the learning abilities of new subjects.

Fig. 1

A Self Organising Map is represented as a rectangular lattice of hexagonal neurons. Each hexagon represents a neuron, and shared edges represent connections between neurons

Fig. 2

Graphical representation of the SOM after 2000 cycles of training. Neurons are represented by the small blue hexagons. The distances between the weight vectors associated with each neuron are colour coded such that light colours represent smaller distances, while dark regions encode for larger ones. As such, sequences of dark coloured connections between neurons are interpreted as borders on the map

Fig. 3

Results from the mapping phase of the SOM. The number of experimental subjects each neuron responds to are displayed within the hexagon representing the neuron