Abstract
Purpose :
Understanding and predicting functional outcomes in intravitreal pharmacotherapy is an important step to personalize therapies and to increase treatment efficiency. Our machine learning approach categorizes future change in BCVA from demographic and clinical data and spectral-domain optical coherence tomography (SD-OCT) images. Analysis of learned models reveals the prediction power of OCT vs. non-OCT biomarkers.
Methods :
Prospective phase III clinical trial data of 2555 patients with four leading exudative macular diseases (s. Table 1) were analyzed. We quantified more than 14,000 monthly acquired OCT scans by automatically segmenting retinal layers and intra- and subretinal fluid, and extracted OCT biomarkers such as fluid location, volume or the average thickness of retinal layers. We applied machine learning with Bayesian Networks to these and to the demographic and clinical data (e.g. age, gender, BCVA, treatment), to predict BCVA changes in two scenarios: First, we predicted them to month 3 (M3), year (Y) 1 and 2 from baseline (BL) visit data and second to Y1 and Y2 from the sequence of data to M3. We also evaluated and compared different models by changing the information presented to them in a 2-fold cross-validation setting.
Results :
Table 1 shows BCVA prediction was more accurate for RVO and DME than for nAMD. Also, vision loss from BL is more reliably predicted than vision gain. By contrast, from M3 onwards vision gain achieved better or similar accuracy. Moreover, our analysis in Figure 1 shows:
● For DME, BCVA measurements and demographic data are more predictive than OCT biomarkers,
● For RVO and AMD, OCT biomarkers of BL scans are important in BCVA prediction and combining them with non-OCT features increases their accuracy, but they lose their importance to the non-OCT features from M3 onwards,
● OCT biomarkers from para- and perifoveal segments increase predictive power in all diseases, but predictions from M3 rely mostly on the foveal segment.
Conclusions :
Prediction of BCVA outcome relies more on OCT biomarkers in RVO and nAMD than in DME. Machine learning identifies disease-specific variables for patient management and can guide the ophthalmologist in predicting clinical outcomes.
This is an abstract that was submitted for the 2018 ARVO Annual Meeting, held in Honolulu, Hawaii, April 29 - May 3, 2018.