Abstract:
The electroretinogram (ERG) is an essential method for diagnosing and monitoring hereditary eye diseases, because it allows an objective assessment of retinal function. However, high-level intrinsic electrical and physiological noise often interferes with the physiological signals of interest, rendering the results difficult, if not impossible, to evaluate, especially when the retinal function is already impoverished.
For several years, nonlinear time series analysis has been applied to developing signal processing methods that rely on transforming a scalar sequence of measurements into a multidimensional vector space, known as "phase space". This transformation, or embedding, allows one to recover implicit information about the system of interest that is not accessible by analysis of the recorded scalar time series alone. Moreover, embeddings form a fundamental starting point for several mathematical-physical noise reduction (NR) algorithms that exploit the signal's nonlinear properties. Therefore, the application of such methods to the analysis of complex systems seems very promising, particularly when nonlinear behaviour is observed, as is the case for the voltage-time series comprising the ERG. Until now, however, the feasibility of applying such methods to the ERG has not been considered.
The aims of this thesis were twofold: first, to investigate if embeddings themselves add meaningfully to the ERG's spectrum of conventional diagnostic parameters; and, second, to compare the performance of the new nonlinear NR methods with the NR techniques applied in conventional ERG diagnostics.
To achieve this, a PC-operated diagnostic system, named CALYPSO, was developed to record ganzfeld ERGs and to analyze them by means of nonlinear time series analysis. The system is capable of creating and displaying embeddings as well as nonlinear filtering. Additionally, it can run the fast Fourier transform (FFT), finite impulse response (FIR) filtering, and standard averaging techniques. Using the system, 30 Hz flicker ERGs, measured according to the ISCEV standard, of 10 normal volunteers and 4 patients suffering from different hereditary retinal diseases were recorded. The scalar data were reconstructed in phase space by a time delay embedding and then analyzed by means of a newly developed topographic approach (Topographic Angles). For the data at hand, the Topographic Angles turned out to be a powerful marker for distinguishing between normal and pathological ERGs, independently of the diagnostic criteria employed in the ISCEV standard. Thus, Topographic Angles may be regarded as a potential addition to the currently available diagnostic methods.
The second aim of this thesis was to evaluate nonlinear methods for recovering ERG signals. Nonlinear Projective Noise Reduction (NNR) is one example of such algorithms. It is independent of the frequency spectrum of the analyzed signal, which sets it apart from conventional methods like FFT or FIR filtering. NNR is therefore useful in eliminating broad-band noise, whose frequency spectrum closely resembles the spectrum of the signal of interest - a common problem in electrophysiology. The NNR algorithm was implemented in CALYPSO and its performance was evaluated by means of normal ERG data, which were contaminated by different types of artificially generated noise. The initial signal-to-noise ratio (SNR), the type of noise (white noise or spectrally matched noise) and the number of algorithm iterations were varied. The results were compared with results obtained by the classical methods of averaging and by FFT frequency filtering. A subsequent quantitative analysis revealed the advantages and disadvantages of each of the three methods. Averaging performs predictably, is stable and provides NR up to 7dB under the tested conditions, independently of the initial noise level. The FFT is very powerful at low initial SNR and provides NR up to 12dB, but produces severe artifacts at high initial SNR. The NNR performs intermediately, providing NR up to 8dB. In principle, it can even reduce noise whose frequency content is identical to that of the uncontaminated ERG data, which neither FFT nor FIR filtering is capable of. An additional bonus was that all three NR schemes may be combined beneficially without incurring severe redundancy effects. In fact, the three methods contribute fairly independent of each other to a total NR of some 15dB, which is much more than each can provide alone.
Given these advantages, it seems very likely that the phase space representation, Topographic Angles and the NNR will contribute to our modern ophthalmalogical diagnostic tool set in the near future. The development of CALYPSO has been the first successful step. Further investigation is now warranted.