DE
Most advanced approach to data mining and predictive modeling
Life sciences and healthcare data are notoriously complex and contain large sets of often interrelated features. Sophisticated modeling techniques are required to handle the high-dimensional, nonlinear dynamics in the data.
Exploris Health has built a unique and autonomous data mining and predictive modeling software package called AI-X-ENGINE, based on a multi-level modeling architecture, which delivers the highest possible quality of classification and prediction by combining, evaluating, and optimizing the use of various methods in an automated self-learning process based on an evolutionary optimization procedure.
The imitation of evolutionary principles in pattern recognition and variable selection enables a sophisticated combination and nesting of different methods from the field of artificial intelligence.
Models are built using a combination of associative mathematical structures like an Artificial Neuronal Network (ANN), particularly Self-Organizing Maps (SOM), or Decision Tree Ensembles and other Ensemble Methods with an Artificial Evolutionary Procedure (AEVOP)3,4. The methods used is best conceived as a heterogeneous adaptive system (HAS)5, employing competitive learning driven by artificial evolution as a basis, and using statistical models for selecting the finally desired model. The ensemble methods and SOM are able to detect autonomously hidden correlational structures, even if they are non-linear. The approach reaches this target using multi-level re-sampling and cross-validation where each pattern goes through a number of stability and performance checks competing with other patterns for survival and only the fittest survive in a final model.
The entire modeling process is highly automated and standardized. No pre-evaluation of data attributes by experts is needed – the software detects hidden relationships in a given data sample (e.g., a clinical study) by its own and evaluates the relevant prognostic factors.
The ability to steer the modeling process in advance in direction of the desired sensitivity/ specificity allows for building models with very low false positive or false negative rates, which is not achievable with standard methods.
Delivering predictive models that can detect and reproduce highly complex non-linear relationships in data of large dimensionality.
The feature selection and dimensionality tasks are performed autonomously. The manual try and error method is replaced with a directed search for the best possible methods and attributes combinations which is performed automatically by the system.
Multi-layer cross-validation ensures the stability of the model derived from our approach even with small datasets.
Provide maximum analysis quality and ensure a controlled and traceable analysis process for diagnostic classification.
