Abstract
Introduction. Artificial neural networks (ANN) may yield incredible results in respect to classification of dithered data where classical statistics fails. We present a ready-to-use computer programme for hormone surge detection based on principles of ANN.
Methods. 15 plasma growth hormone lime series (8 to 11 hours, one value every 20 min) of different patients were visually evaluated by an expert endocrinologist (JHB), and ‘true’ pulses marked without comment. Next, facts were automatically generated out of those preevaluated data sets of not necessarily equal length using the ‘shilling window’ technique, and were presented to a PC-based ANN (back-propagation, 2-layer topology, 11 inputs, 7 hidden-, I output neuron, language ‘C’) in random order and repeatedly. The ability to self-organize inherent to the ANN enabled it to associate output (puls/no puls) with general input patterns (‘learning’). RESULTS. After a training phase of 400 epochs (20 minutes, 386-AT with coprocessor), the errors F1/F2 had decreased from 175 to 3.4 and 10 to 0.26, respectively. Interpretation of all facts was correct. Although the number of facts was small, with independent test data, the satisfying system's capability to generalize could he demonstated. CONCLUSION. The system enables laboratory staff not specially trained i) to train their local PC-based ANN with prccvaluated hormone time series, and thereafter ii) to routinely perform hormone surge detection. The depersonalized expert knowledge becomes consistent and reproducible without the need to explicitely define pulse criteria. In summary, due to the neural network's self-organizing and learning properties, the programme is a handy tool to evaluate all kinds of episodic hormone secretion.
Article PDF
Author information
Authors and Affiliations
Rights and permissions
About this article
Cite this article
Michel, E., Volmer, R., Brämswig, J. et al. NEURAL NETWORK BASED ANALYSIS OF EPISODIC HORMONE SECRETION FOR CLINICAL ROUTINE. Pediatr Res 35, 263 (1994). https://doi.org/10.1203/00006450-199402000-00053
Issue Date:
DOI: https://doi.org/10.1203/00006450-199402000-00053