PURPOSE: Pattern electroretinograms (PERGs) have inherently low signal-to-noise ratios and can be difficult to detect when degraded by pathology or noise. We compare an objective system for automated PERG analysis with expert human interpretation in children with optic nerve hypoplasia (ONH) with PERGs ranging from clear to undetectable. METHODS: PERGs were recorded uniocularly with chloral hydrate sedation in children with ONH (aged 3.5-35 months). Stimuli were reversing checks of four sizes focused using an optical system incorporating the cycloplegic refraction. Forty PERG records were analysed; 20 selected at random and 20 from eyes with good vision (fellow eyes or eyes with mild ONH) from over 300 records. Two experts identified P50 and N95 of the PERGs after manually deleting trials with movement artefact, slow-wave EEG (4-8 Hz) or other noise from raw data for 150 check reversals. The automated system first identified present/not-present responses using a magnitude-squared coherence criterion and then, for responses confirmed as present, estimated the P50 and N95 cardinal positions as the turning points in local third-order polynomials fitted in the -3 dB bandwidth [0.25 … 45] Hz. Confidence limits were estimated from bootstrap re-sampling with replacement. The automated system uses an interactive Internet-available webpage tool (see http://clinengnhs.liv.ac.uk/esp_perg_1.htm). RESULTS: The automated system detected 28 PERG signals above the noise level (p ≤ 0.05 for H0). Good subjective quality ratings were indicative of significant PERGs; however, poor subjective quality did not necessarily predict non-significant signals. P50 and N95 implicit times showed good agreement between the two experts and between experts and the automated system. For the N95 amplitude measured to P50, the experts differed by an average of 13% consistent with differing interpretations of peaks within noise, while the automated amplitude measure was highly correlated with the expert measures but was proportionally larger. Trial-by-trial review of these data required approximately 6.5 h for each human expert, while automated data processing required <4 min, excluding overheads relating to data transfer. CONCLUSIONS: An automated computer system for PERG analysis, using a panel of signal processing and statistical techniques, provides objective present/not-present detection and cursor positioning with explicit confidence intervals. The system achieves, within an efficient and robust statistical framework, estimates of P50 and N95 amplitudes and implicit times similar to those of clinical experts.
PURPOSE: Pattern electroretinograms (PERGs) have inherently low signal-to-noise ratios and can be difficult to detect when degraded by pathology or noise. We compare an objective system for automated PERG analysis with expert human interpretation in children with optic nerve hypoplasia (ONH) with PERGs ranging from clear to undetectable. METHODS: PERGs were recorded uniocularly with chloral hydrate sedation in children with ONH (aged 3.5-35 months). Stimuli were reversing checks of four sizes focused using an optical system incorporating the cycloplegic refraction. Forty PERG records were analysed; 20 selected at random and 20 from eyes with good vision (fellow eyes or eyes with mild ONH) from over 300 records. Two experts identified P50 and N95 of the PERGs after manually deleting trials with movement artefact, slow-wave EEG (4-8 Hz) or other noise from raw data for 150 check reversals. The automated system first identified present/not-present responses using a magnitude-squared coherence criterion and then, for responses confirmed as present, estimated the P50 and N95 cardinal positions as the turning points in local third-order polynomials fitted in the -3 dB bandwidth [0.25 … 45] Hz. Confidence limits were estimated from bootstrap re-sampling with replacement. The automated system uses an interactive Internet-available webpage tool (see http://clinengnhs.liv.ac.uk/esp_perg_1.htm). RESULTS: The automated system detected 28 PERG signals above the noise level (p ≤ 0.05 for H0). Good subjective quality ratings were indicative of significant PERGs; however, poor subjective quality did not necessarily predict non-significant signals. P50 and N95 implicit times showed good agreement between the two experts and between experts and the automated system. For the N95 amplitude measured to P50, the experts differed by an average of 13% consistent with differing interpretations of peaks within noise, while the automated amplitude measure was highly correlated with the expert measures but was proportionally larger. Trial-by-trial review of these data required approximately 6.5 h for each human expert, while automated data processing required <4 min, excluding overheads relating to data transfer. CONCLUSIONS: An automated computer system for PERG analysis, using a panel of signal processing and statistical techniques, provides objective present/not-present detection and cursor positioning with explicit confidence intervals. The system achieves, within an efficient and robust statistical framework, estimates of P50 and N95 amplitudes and implicit times similar to those of clinical experts.
Authors: Daphne McCulloch; Pamela Garcia-Filion; Cassandra Fink; Anthony C Fisher; Antonio Eleuteri; Mark S Borchert Journal: Doc Ophthalmol Date: 2017-08-09 Impact factor: 2.379
Authors: Ruth Hamilton; Michael Bach; Sven P Heinrich; Michael B Hoffmann; J Vernon Odom; Daphne L McCulloch; Dorothy A Thompson Journal: Doc Ophthalmol Date: 2020-06-02 Impact factor: 2.379