eyes that had not yet shown clinical retinopathy and to explore whether abnormal neural function measures might be predictive of new retinopathy development.

Our studies over the 4–5 years that followed confirmed the initial promise of this measure, and we now know that the neural latency abnormalities (mfERG delays) observed in the earlier studies are not only present prior to retinopathy onset [19, 20] and correlated with the severity of the retinopathy at the local site of the retinal vascular signs of retinopathy [19–21], but are also predictive (precedes) retinopathy onset at locations in eyes that already have some retinopathy [20, 22–24]. Our longitudinal studies over 1, 2, and 3 years have shown that predictive models based on mfERG delays revealed remarkable potential clinical and research tools with high sensitivity and specificity [20, 23, 24]. In confirmation studies, the high sensitivity (prediction of retinopathy onset in a specific location) and specificity (prediction of normal retina at specific locations) remain high [24]. One of our recent publications also reveals that the mfERG measures are predictive of the onset of retinopathy in eyes that had no prior retinopathy [25].

These research results with early stage emergence of neural dysfunction measures in the retina are in striking contrast to the natural history of change in visual acuity. Visual acuity loss occurs many years after retinopathy appears, and then only with severe retinopathic events or edema that impact the fovea. By that time, the vascular events are very apparent to the clinician. So, as a functional outcome measure, visual acuity is primarily useful as a measure of success in slowing late-stage retinopathy, or for assessing the impact of treatments applied at that late stage. It is not useful to signal imminent retinal problems, early retinopathy progression, or the efficacy of any preventative treatments. In contrast, the implicit time (delay) measure of the mfERG has emerged as an exciting future clinical tool in the management of patients at earlier stages and for the exploration of new candidate treatments and interventions. With it, we have produced formal predictive models. It is the critical component of predictive models of retinopathy onset over a relatively short time frame and, as such, is an obvious candidate as an outcome measure for relatively brief clinical trials of proposed pharmaceutical preventatives at the earliest stages of diabetic retinal complications.

Predictive Models of Visible Retinopathy Onset at Specific Locations

Using multivariate logistic regression techniques, we formulated models incorporating mfERG IT and risk factors such as duration of diabetes and blood glucose control that predict the development of retinopathy in new retinal locations with high sensitivity and specificity (approx. 80–90%) [20, 23, 24]. Recently, we formulated another multifactor model, based on mfERG IT, that predicts the initial clinical onset of diabetic retinopathy [25].

HOW IS THE MFERG MEASURED AND WHAT IS IT MEASURING?

So, what is the mfERG and how is it actually measured? The mfERG is a noninvasive technique for measuring neural function in up to hundreds of contiguous retinal areas within the central retina [26, 27]. The implicit time (IT) of the P1 component of the local mfERG response waveform is a highly reproducible and sensitive indicator of neural

function in the retina. Figure 1 provides a brief overview of the stimulus and response outputs across the central 45° of human retina [18–30].

Briefly, 103 local retinal responses to 200 cd/m2 flash stimulation (actually, first-order response kernels) are recorded from the central ~45° of the retina during an ~8 min session using a 75 Hz frame rate and 10–100 Hz filtering. The responses are recorded using a bipolar contact lens electrode, and a ground electrode is clipped to the right earlobe. Fixation is monitored using an in-line infrared video camera. The session is broken into 16 segments for subject comfort.

The first prominent positive peak (P1) of the mfERG response (Fig. 1D) that our group has investigated is the easiest to measure, and the implicit time measure of P1 is far less variable than the amplitude measure (one-tenth of the coefficient of variation of the amplitude measure in healthy control subjects) [20].

Where Are These Neural Signals Generated in the Retina?

It is generally believed that mfERG IT delay, in the absence of reduced response amplitude, reflects abnormality of the outer plexiform layer and bipolar cells, as it does for the conventional full-field flash ERG. The P1 component of the mfERG waveform, from which we measure mfERG IT, is generated primarily by the opposing electrical polarities of the ON and OFF bipolar cell responses in the middle layers of the retina [31–33]. The retina is particularly susceptible to the early pathological vascular changes associated with type 2 diabetes because of its high metabolic demand, minimal retinal vascular supply, and low oxygen tension of the inner retinal layers [8]. It has been proposed that mfERG IT delays in the absence of mfERG response amplitude reductions represent the effects of reduced perfusion and resulting hypoxia/ischemia [19, 20, 22, 23, 30, 34]. Recently, more direct evidence supporting this view has been reported. In diabetic patients with enlarged foveal avascular zones, the area of the vascular-free zone has been shown to be correlated with increasing mfERG IT delay, but not mfERG amplitude reduction, in and adjacent to the fovea [34].

Some Key Results

Before highlighting the evolution of our predictive models, since 2004, it is illustrative to look at a single patient example of the way in which the local mfERG implicit time delay predicted subsequent retinopathy in a patient (Fig. 2).

In one of our first publications, we reported the sensitivity and specificity of the mfERG implicit time as part of a “one-year” predictive model. It certainly included what we later learned were both retinopathy that was transient and retinopathy that was likely to be persistent. Based on our data then, primarily for study patients with mild diabetic retinopathy, we found relatively high sensitivity (86%) and specificity (84%) [23].

This quantitative model was the first to make predictions of diabetic retinopathy lesions in discrete retinal areas. The study involved only one follow-up visit and thus could not examine whether the lesions that were evidenced were transient or sustained in nature. More recently, our review included new data that extended the study by Han et al. [23] for another year [20].

Two years later, we reported on a 3-year prediction model with similarly high sensitivity and specificity for patients who already had some retinopathy [24]. Eighteen

Fig. 2. Shows an example of the predictive power of the mfERG IT. The baseline mfERGs are shown in (A). At baseline, this patient had no retinopathy. The mfERG implicit time was, however, abnormally delayed (P < 0.023) in many of the 103 locations (red hexagons in (B)) and many of the 35 retinal zones used for analysis (red patches in (D)). On follow-up 1 year later, new patches of retinopathy and edema had developed, as indicated in the fundus photograph (C) and as black dots (D). As can be seen in (D), four of the five new lesions are associated with significantly delayed mfERG IT one year earlier, and the fifth lesion is very close to a delayed zone. (Fig. 2 from Bearse et al. [20]).

diabetic patients were examined at baseline and at three annual follow-ups. Again, 35 retinal zones were constructed from the 103-element stimulus array, and each zone was assigned the maximum IT z-score within it based on 30 age-similar control subjects. Logistic regression was used to investigate the development of retinopathy in relation to baseline mfERG IT delays and additional diabetic health variables. Again, receiver operating characteristic (ROC) curves were used to evaluate the models.

Fig. 3. (from Ng et al. [24]) ROC curves for the multivariate model (right) that predicts recurring retinopathy over the course of 3 years in a local retinal area. The area under the ROC curve (AUC) of 0.95 provides an overall measure of the model’s discrimination accuracy (95%). Even a model containing only the mfERG implicit time, and no other factors (left), provided surprisingly good sensitivity (84%) and specificity (73%) along with good discrimination (AUC = 0.83) [24].

Here, we were interested in the prediction of persistent retinopathy onset at two successive annual visits. We looked separately at what we had learned was a common occurrence of transient initial retinopathy. A retinal area that shows retinopathy lesions over a longer period represents more significant pathophysiological alterations—increased vascular permeability and hypoxia. We argue that these areas are clinically more important than transient retinal lesions. (It is well known that the very earliest clinical signs of diabetic retinopathy wax and wane. For example, Hellstedt and Immonen reported that over a 2-year period, 52% of microaneurysms show spontaneous resolution [35]. )

Retinopathy developed in 77 of the 1,208 retinal zones of which 25 retinal zones had recurring retinopathy. The multivariate analyses showed baseline mfERG IT, duration of diabetes, and blood glucose concentration as the most important predictors of recurring retinopathy but were not at all predictive of transient retinopathy. ROC curves revealed sensitivity of 88% and specificity of 98% for the recurring retinopathy we were interested in (see Fig. 3). A tenfold cross-validation confirmed the high sensitivity and specificity of the model.

In a recent publication, we report on the onset of diabetic retinopathy in a study group of patients with diabetes but no clinically visible signs of retinopathy [25]. Again, the predictive multivariate models incorporating mfERG IT delay and other risk variables showed excellent ability to predict the onset of retinopathy with high sensitivity and specificity. Seventy-eight eyes from 41 diabetes patients were tested annually for several years. The presence or absence of DR at the last study visit was the outcome measure, and measurements of risk factors from the previous visit were used for prediction. Nearly 40% of the 78 eyes developed retinopathy for a total of 80 of 2,730 retinal zones. In short, mfERG IT was again a good predictor of diabetic retinopathy onset,