Augmentation of telemedicine post-operative follow-up after oculofacial plastic surgery with a self-guided patient tool

ABSTRACT Purpose This study evaluates a web-based tool designed to augment telemedicine post-operative visits after periocular surgery. Methods Adult, English-speaking patients undergoing periocular surgery with telemedicine follow-up were studied prospectively in this interventional case series. Participants submitted visual acuity measurements and photographs via a web-based tool prior to routine telemedicine post-operative visits. An after-visit survey assessed patient perceptions. Surgeons rated photographs and live video for quality and blurriness; external raters also evaluated photographs. Images were analyzed for facial centration, resolution, and algorithmically detected blur. Complications were recorded and graded for severity and relation to telemedicine. Results Seventy-nine patients were recruited. Surgeons requested an in-person assessment for six patients (7.6%) due to inadequate evaluation by telemedicine. Surgeons rated patient-provided photographs to be of higher quality than live video at the time of the post-operative visit (p < 0.001). Image blur and resolution had moderate and weak correlation with photograph quality, respectively. A photograph blur detection algorithm demonstrated sensitivity of 85.5% and specificity of 75.1%. One patient experienced a wound dehiscence with a possible relationship to inadequate evaluation during telemedicine follow-up. Patients rated the telemedicine experience and their comfort with the structure of the visit highly. Conclusions Augmented telemedicine follow-up after oculofacial plastic surgery is associated with high patient satisfaction, rare conversion to clinic evaluation, and few related post-operative complications. Automated detection of image resolution and blur may play a role in screening photographs for subsequent iterations of the web-based tool.


Introduction
Coronavirus disease of 2019 (COVID-19) caused a dramatic change in healthcare delivery in the United States. Prior to this public health emergency, telemedicine played a limited role in ophthalmic care, including such applications as diabetic retinopathy screening, lowrisk glaucoma follow-up, and remote consultation of incidental findings. 1 Social distancing and shelter-inplace mandates led to limitations in evaluating patients in a traditional clinic setting, which in turn resulted in rapid adoption of telemedicine services across many healthcare disciplines. 2,3 The authors' institution mandated that telemedicine services be used whenever in-person contact was not considered essential by the treating physician. This mandate posed significant examination and diagnostic challenges for many disciplines in ophthalmology, but it was felt that the subspecialty of oculofacial plastic surgery was uniquely positioned to adopt this technological shift more readily, due to to a lower dependence on specialized imaging and slit-lamp examination. Thus, the authors converted the first postoperative visits after many kinds of oculofacial plastic surgeries to remote, audio-visual encounters. This endeavor was quickly noted to have some shortcomings, including a lack of visual acuity assessment and variable quality of the video component for judging surgical outcomes. As such, the authors devised a supplemental telemedicine protocol involving a pre-visit web portal to guide patients through the process of measuring visual acuity, documenting a basic history, and obtaining high-quality photographs. Herein, the authors examine the utility of the web-based self-guided patient tool to augment postoperative telemedicine visits among patients having oculofacial plastic surgery during the COVID-19 pandemic.

Methods
This prospective, interventional case series was conducted at the Department of Ophthalmology, University of California, San Francisco, USA, with approval from the Institutional Review Board (IRB) (#20-31122). Participants were enrolled between June 2020 and June 2021, with follow-up through December 2021. This study adhered to the tenets of the Declaration of Helsinki as amended in 2013 and was compliant with the Health Insurance Portability and Accountability Act (HIPAA). Informed consent was obtained verbally, provided in written form electronically on the study website, and confirmed implicitly by patients' submission of survey data; the requirement for written consent was waived by the IRB due to the public health crisis and minimal risk nature of the study. Written informed consent was obtained for the publication of one participant's photographs.

Telemedicine protocol
Patients at the authors' institution are routinely seen at a 1-week and 3-month interval after most types of oculofacial plastic surgery. During the public health crisis, many patients were transitioned to telemedicine evaluation for their 1-week postoperative visit. This was done through a secure, standard definition (720p) audio-visual connection (Zoom, Zoom Video Communications Inc., San Jose, California). A telephone visit was used if technical difficulties were encountered. In-person examination was available to patients for needs deemed urgent by either the patient or the surgeon. Post-operative discharge instructions included standard language describing warning signs to alert patients of scenarios in which to contact their surgeon.
Study participants completed a web-based study activity prior to their telemedicine visits. Postoperative discharge instructions included both a Uniform Resource Locator (URL) and a Quick Response (QR) code linking to a web-based Research Electronic Data Capture (REDCap) survey. Study data were then collected and managed using REDCap electronic data capture tools hosted at UCSF. REDCap is a secure, web-based software platform designed to support data capture for research studies. 4,5 The pre-visit tool (Supplemental Figure S1) contained three components. First, step-by-step instructions were provided for visual acuity measurement, using either a paper Snellen chart or a clinically validated smartphone application (Verana Vision Test, Verana Health Inc., San Francisco, California). 6,7 Second, a selfguided history was obtained including pain levels, subjective vision, medication problems, and patient concerns. Finally, guidance was provided for facial photography to produce high-quality, standardized images of participants' surgical sites. This guidance included instructions to focus the image, to have a companion assist with using the higher quality rear smartphone camera, to hold the camera about 18-24 inches away to minimize fisheye distortion, and to utilize flash unless lighting were optimal. Participants were instructed to obtain a minimum of three photographs: primary gaze, up-gaze, and eyelid closure. An example of each photograph was shown. Participants were also provided an option to submit additional photographs of areas of concern.

Eligibility criteria
Subjects were recruited from the population of patients scheduled for telemedicine follow-up after oculofacial plastic surgery. Clinical eligibility for telemedicine was a joint decision by the surgeon and patient based on: technological feasibility, comfort and preference, surgical risk/acuity, and intraoperative complications. Generally, higher risk surgical procedures such as orbitotomy or patients experiencing intraoperative complications had in-person follow-up unless the patient strongly preferred telemedicine follow-up, whereas most patients undergoing uncomplicated, routine surgeries were encouraged to utilize telemedical care.
Among this telemedicine population, eligibility was assessed for study participation. Eligible patients were 18 years or older, spoke English as their primary language, and reported comfort with navigating an online interface.

Multimedia analysis
The operating surgeon (BJW, RCK, or MRV) and assisting junior surgeon (DCA) assigned a quality score to both photographs and video during the first postoperative visit. Quality was defined as the subjective usefulness for evaluating post-operative appearance, including the surgical outcome and potential complications, and was scaled from 1 (so poor as to be useless) to 10 (comparable to in-person evaluation). For the photograph in primary gaze, a binary assessment of the subjective presence of blurriness was also recorded. Two junior surgeons (LDS, MA) who were masked except to the type of surgery also rated photographs for both quality and blurriness.
The primary gaze photographs were quantitatively evaluated for three features to identify their respective contributions to image quality. First, image resolution in megapixels (MP) was calculated based on image dimensions obtained from the file property function of the Windows operating system (Windows 10 Professional; Microsoft Inc., Redmond, Washington). Second, centration of each participant's face was measured, with the midpoint between the two pupils (i.e. over the nasal bridge) as the reference point. Horizontal and vertical centration were calculated as fractions of the reference point's position relative to the respective dimensions. The mean of these two fractions was used for overall centration. Facial decentration was the absolute difference between actual and ideal (0.50) centration.
Third, image blurriness was measured based on theory described by Liu et al. 8 Transformation of images from the spatial domain to the frequency domain is a means of facilitating edge and blur detection using computer vision. 9 Blurry regions of images have relatively little change from pixel to pixel, and therefore are represented with low frequencies in the frequency domain. In contrast, sharp regions of images, such as a well-focused transition from a patient's face to the background, have relatively high change from pixel to pixel, and thus are represented with high frequencies in the frequency domain. An image with generally more high-frequency regions is less blurry than an image with more low-frequency regions. An open-source algorithm implementing this concept was used to quantify the blur of each primary gaze photograph. 10 This process is illustrated in Figure 1. Images were converted to grayscale and scaled to a standardized width (500px; Figure 1a). A Fast Fourier Transformation (FFT) was carried out to convert each image to the frequency domain, and the result was shifted to centralize the zero frequency component. Next, a high-pass filter was applied in a central, 75 × 75 pixel square, selectively leaving the high frequencies that corresponded to the edges and details that make up a sharp photograph (e.g. a facial outline; Figure 1b). Then, an inverse shift and inverse FFT were performed to reconstruct the image (Figure 1c). The magnitude spectrum of the image was computed and the mean was calculated. The mean magnitude corresponds to the relative presence of sharp transitions in the image (low-frequency data having been filtered out), and thus the degree of blur; a higher mean magnitude indicates a sharper image, and vice versa. A threshold value is required to classify the mean as blurry or sharp; this can vary based on factors such as the lighting, subject matter, and equipment. Therefore, the subjective blur assessments described above were used to identify a threshold for this application of the algorithm, in which there is a reasonably standardized, real-world setting. The majority opinion (i.e. ≥2 of 3 raters) was used as ground truth for each image.

Complications
Chart review was carried out for a minimum of 3 months after surgery to identify complications as reported by the surgeon, impromptu clinic visits, return to the operating room, and hospitalization or emergency department evaluation for issues related to surgery.
Complications were defined as undesirable and unexpected events. Complications were considered major if there were permanent functional abnormalities (e.g. vision loss, diplopia), major surgery required for revision/correction, or medical intervention requiring hospitalization. Complications were considered moderate if there were transient functional abnormalities, minor/ office-based surgery required for revision/correction, or medical intervention not requiring hospitalization. Complications were considered minor if no significant intervention was required. Thus, complications of moderate or major severity are most consistent with the typical use of the term; minor complications were included nonetheless in this study to err on the side of over-reporting any adverse events that might be associated with telemedicine care.
Complications were further assigned an estimate of their relationship to telemedicine follow-up. "None" was assigned when a relationship could be ruled out due to identification of the complication before or during the telemedicine visit. "Inadequate evaluation" was assigned when a complication was recognized during the telemedicine visit but required in-person evaluation for definitive diagnosis and management. To aid with these distinctions, surgeons documented the circumstances of complications identified during the telemedicine visit. For complications occurring after the telemedicine visit, "Possible" or "Unlikely" were assigned to indicate the suspected probability that in-person examination at the first visit could have identified an abnormality earlier, leading to prophylactic action and a different outcome. These indefinite terms were chosen in recognition of the difficulty in establishing causation, and assessment of the relationship was based on the natural history of the complication.

Participant after-visit survey
Study participants were asked to complete an optional, non-clinical REDCap survey after their telemedicine visits (Supplemental Figure S2) using Likert-type scales. The construction of the survey is described in Supplemental Methods 1.

Statistical analysis
Data analysis was performed using R (R Core Team, 2022). Means were compared with the t-test for unpaired data, and the paired t-test for intrasubject data. The Krippendorff alpha was calculated to evaluate inter-rater reliability for photograph quality. 11 The pROC R package was used to generate a receiver operating characteristic (ROC) curve for photograph blur analysis. 12 Multiple logistic regression was used to explore predictive factors for multimedia quality and patient-reported outcomes. Vision was converted to logarithm of the minimum angle of resolution (logMAR) notation for the purposes of analysis. In order to compare paper-versus smartphone application-based visual acuity measurement, the difference between the tool and clinic visual acuity measurements were first calculated for each patient. The mean of the differences for patients using the smartphone application and those using the paper chart were then compared with a t-test. A P-value of <0.05 was considered significant.

Results
During the study enrollment period, 498 patients underwent surgery with one of the three study surgeons at the university hospital. Among these, 122 patients were eligible for participation after study criteria were applied, and 79 (64.7%) elected to participate. Most study enrollment occurred in the first 6 months (67%), after which COVID-19 vaccinations became available to the general population, leading to relaxation of public health restrictions and a decrease in the proportion of patients opting for telemedicine follow-up. The median age of participating patients was 62 years (interquartile range [IQR] 21.0, range 19-89). The median follow-up time was 97 days (IQR 91.5, range 4-509), with 62 patients (78.4%) having at least 3 months of postoperative follow-up. The types of surgeries performed are summarized in Table 1.
All patients submitted pre-visit photographs, while 78 (98.7%) participated in a video visit and one (1.3%) attended a telephone visit due to technological difficulties. Video and photograph quality scores are shown in Table 2. On average, surgeons assigned a higher quality score to photographs than live videos for evaluating postoperative recovery (mean difference 1.0; 95% CI [0.7, 1.4]; p < 0.001). Each decade of age was associated with a 0.47 lower odds of subjective photograph quality in the highest 75 th percentile (95% CI [0.44, 0.51]; p = 0.032), while no association was identified for live video quality (Supplemental Table S1). Photographs with resolution lower than that of the maximum video resolution (0.9 MP; n = 8) received lower subjective photograph quality scores than photographs with higher resolution (6.8 vs. 7.8; 95% CI [−3.0, 0.9]; p = 0.248) without reaching significance. Krippendorff's alpha for the three raters of photograph quality was 0.647, indicating moderate agreement. 11 Quantitative analysis of primary gaze photographs demonstrated mean facial centration of 48.2% and resolution of 8.1MP. Subjective blur was identified in 16 photographs (20.3%) based on a majority opinion of the three raters. These features were plotted against photograph quality and are shown in Figure 2. The mean magnitude values for each photograph as determined by the blur detection algorithm ranged from −4.9 to 31.2 dB (lower is blurrier), with a sample mean of 15.4 dB (Standard deviation [SD] = 7.5). The ROC curve for the algorithm's blur score with subjective blur as ground truth is shown in Figure 3. The area under the curve (AUC) was 0.88 [95% CI 0.73, 0.96], indicating excellent discrimination by the algorithm. 13 A threshold value of 11.3 dB for the algorithm score achieved the highest sum of sensitivity and specificity at 85.5% and 75.1%, respectively.
Clinic visual acuity measurements were available from 79 patients (100%) at the pre-operative visit and 59 (74.6%) at the final post-operative visit. Self-guided, telemedicine visual acuity measurements were available from 79 patients (100%) for the early post-operative visit. Mean visual acuity from the telemedicine visit was slightly worse than mean visual acuity at both the pre-operative and final post-operative clinic visits by approximately one line on a Snellen chart (Table 3). In contrast, initial and final clinic visual acuity were not significantly different.
Fifty-seven patients (72.2%) opted to measure visual acuity with a paper chart, while 22 (27.8%) chose to measure with the smartphone application. Those using the application deviated less from their preoperative visual acuity than those using the paper chart (0.04 vs.    telemedicine experience, and comfort with the visit structure. Age and multimedia quality were not statistically significant predictors of patient-reported visit comfort (Supplemental Table S1). When asked how they would prefer a hypothetical, future first postoperative visit to occur, 16 (59.3%) chose a video encounter with a pre-visit tool including photographs, 7 (25.9%) chose an in-person evaluation, 2 (7.4%) chose a video encounter alone, and 2 (7.4%) chose a telephone encounter with a pre-visit tool including photographs. Age, multimedia quality, duration of clinic and telemedicine visits, and distance traveled to clinic were not found to be statistically significant predictors of future visit preference (Supplemental Table S1). Telemedicine encounters had a self-reported mean duration of 30.    Potential and confirmed complications are summarized in Supplemental Table S2. Six patients (7.6%) had abnormalities noted during the telemedicine visit requiring conversion to an in-person evaluation. Half of these had a reassuring examination in clinic, while the other half had moderate complications. Three patients (3.8%) had complications identified during the telemedicine visit without the need for earlier in-person evaluation, including two patients with over-correction after blepharoptosis repair and one with a small wound dehiscence after wedge resection. Five patients (6.3%) called with concerns before their routine post-operative month three visit and were evaluated in-person earlier.

Discussion
In this study, the authors employed a patient-driven web-based tool to augment telemedicine post-operative visits with visual acuity measurements and photographs, while evaluating complications and patient perceptions of the encounter. Patients reported a high degree of satisfaction with the telemedicine experience. This finding is consistent with that of research among veterans undergoing general surgery procedures with telephone follow-up 14,15 and of reports on telemedicine for nonperioperative care. 16,17 Indeed, the majority of patients (75%) surveyed in the present study indicated that they would prefer another telemedicine visit after a hypothetical future surgery. These results are similar to the results of a survey among oculofacial plastic surgery patients seen for both new and return visits via telemedicine, in which 62% preferred that a hypothetical future visit be remote rather than in-person. 18 The convenience and efficiency of telemedicine visits is likely an important factor behind patients' satisfaction. Patients in this study reported that telemedicine follow-up saved an average of two hours and 70 miles of travel when compared with an in-person visit. The cost savings in travel alone can be estimated at $39.48 based on an expense of 56 cents per mile per the Internal Revenue System. 19 An even higher average cost savings of $888 per visit was found in an economic analysis of telemedicine at another tertiary care center, when considering the value of time and the cost of the lodgings required by many patients; notably, the average savings was still $256 when excluding patients requiring accomodations. 20 Though respecting the time and money of patients is important, it is imperative that the quality of care not be compromised. The findings in this study suggest that telemedicine offers a safe mechanism for oculofacial plastic surgeons to evaluate recovery. Surgeons graded both photographs and live video to be of generally high quality for evaluating postoperative appearance, and converted few visits to in-person examination for better assessment. Those converted visits in which an abnormal finding was identified all involved slit-lamp examination, which is difficult to emulate with photo or live video using current technology. The rate of conversion supports a previous survey of oculofacial plastic surgeons in which 60% of respondents reported "very rarely" converting visits to in-person and 30% reported converting "less than half" of visits. 21 Surgeons were also able to evaluate vision using self-measured visual acuity. These measurements were found to be within one Snellen line of clinic measurements, suggesting that the data provided an accurate approximation for clinical use. Finally, the rate of surgical complications appeared to be within the expectations of prior reports.
All major complications involved reoperations. Reoperation rates of 6.9-8.7% have been reported for blepharoptosis repair 22,23 and 3.2-5.2% for upper blepharoplasty, 24 which are similar to those identified in this study. The failure rate of dacryocystorhinostomy was 18.4% in a large review, 25 versus 25% in the present study; notably, the sample size was small for dacryocystorhinostomy, and thus the high reoperation rate was based on a single case. Reoperation after orbitotomy occurred due to concern for recurrent neoplasm greater than 6 months after surgery, without apparent relationship to telemedicine follow-up; additionally, the sample size for orbitotomy was limited. Lower-grade complications also appeared in line with norms for periocular surgery: wound dehiscence has been reported among 1.4-3.5%, [26][27][28] infection among 0.04-1.2%, [29][30][31][32][33] pyogenic granuloma among 1.6%, 28 and suture inclusion cysts/milia among 2-17%. 34 One moderate complication, wound dehiscence after the telemedicine visit, was potentially related to telemedicine since it was unclear if a more detailed examination could have identified an earlier dehiscence. It is notable that the photographs of this patient were rated highly in quality and did not appear to indicate any abnormality. Otherwise, complications occurred weeks to months after the telemedicine visit and involved issues that were unlikely to be actionable or present at the first visit, such as blepharoptosis undercorrection (which the authors allow to heal) or irritation from abnormally persistent absorbable sutures (which would not be removed when asymptomatic at the first visit). Notably, some surgeons advocate for early adjustment after blepharotposis repair, [35][36][37][38] which raises the important point that telemedicine follow-up must be tailored to individual practice patterns.
The web-based tool utilized in this study provided proof-of-concept for augmentation of the standard video encounter, and may be used to build more complex and user-friendly tools. As previously mentioned, self-measured visual acuity prior to the telemedicine visit was found to be within one Snellen line of clinic visits before and after surgery. This difference could represent a small error in measurement, or a true, mild decrease in vision from ointment, swelling, and other post-operative factors. Regardless, it appears to provide an excellent approximation of vision, which is an important aspect of postoperative ophthalmic care.
Patients were offered the opportunity to use either a smartphone application or a paper chart to measure visual acuity. Interestingly, those using the smartphone application had mean vision measurements closer to clinic measurements. The apparent superiority of application-based measurement may be attributable to error in estimating the appropriate testing distance for the paper chart. The application instructs users to hold the phone at a comfortable reading distance, whereas the paper chart requires a 10-foot distance. The former is easier to implement with intuition rather than measurement, and indeed, a previous study found that people naturally hold smartphones at approximately the 14inch distance that is ideal for a near chart. 39 Thus, an application may be a preferable mode of vision measurement for telemedicine, though notably self-selection of the testing modality potentially introduces bias.
Photographs were an essential contribution of the web-based tool. Surgeons graded the photographs to be of higher quality than corresponding live video for judging postoperative recovery, which is likely related to the inherent difference in image quality between video and photo. The video software used at the authors' institution has a resolution of 720p, or 0.9 MP. This resolution is reduced, sometimes substantially, to maintain fluid video when there is suboptimal hardware or internet connection, and thus it represents a maximum resolution. On the other hand, photographs in this study had a mean resolution of 8.1 MP.
In order to facilitate more effective future applications for collecting patient photographs, the authors explored photograph features associated with higher surgeon-reported quality. Facial centration was generally good in this study, perhaps owing to the detailed instructions and example images guiding patients. This feature was considered because the authors suspected that the autofocus feature of digital cameras might not perform as well for decentered faces. However, there was only a weak correlation between facial decentration and lower photograph quality. Extreme cases did appear to have lower quality, likely due to worse focus or partial capture of the face. There is limited consensus regarding the minimum resolution required for clinical photography; some authors have suggested values ranging from 1.5 to 12 MP, [40][41][42] but these have been based on expert opinion rather than quantitative analysis. The quantitative findings of the present study suggest there is only weak correlation between resolution and subjective image quality in the range produced by modern smartphone cameras; notably, few photographs had values less than 1.5 MP, and there did appear to be a lower mean quality when photograph resolution was at or below that of maximal video resolution (0.9 MP). Not surprisingly, the most influential factor in photo quality appeared to be blurriness.
A future tool could implement the findings of this study to provide a more effective platform for clinical use. Patient guidance and example photos appear sufficient to address facial decentration. The automated detection of low-resolution images is trivial; based on the data in this study, the threshold could be set fairly low, perhaps at 0.9 MP, so that the resolution at least exceeds the upcoming video call. The blur detection algorithm had an excellent AUC when using this study's customized threshold and could be applied to photographs at the time of upload. Patients could then be prompted to attempt repeat imaging based on resolution and blur.
This study has several limitations. The population was pre-selected from patients who had already agreed to have telemedicine follow-up, which may have biased the sample toward those who were more facile with technology. The study did not collect data on non-participants, which limited evaluation of factors associated with the choice to opt against telemedicine follow-up. Notably, fewer patients opted for telemedicine follow-up as public health restrictions were relaxed later in the study, indicating some baseline reticence toward the concept. The high satisfaction of study participants suggests that those patients passing beyond the initial reticence have a good experience in practice, but there is selection bias among the patients who followed through with telemedicine care. Evaluating the causal relationship between telemedicine and complications was limited to surgeon judgment in the absence of a control group; the authors had planned a randomized study prior to the public health emergency (NCT04235803), but randomization to inperson care was subsequently considered unethical. The threshold value for the blur detection algorithm should be subjected to a validation dataset before routine clinical usage; this could be done on a "live" basis by building a clinical platform without blur detection and rating the initial submissions prior to implementing the algorithm. Finally, telemedicine follow-ups after more complex surgeries, such as orbitotomy, were few in number and were biased toward patients with strong telemedicine preference. While this group did not appear to be adversely affected by telemedicine follow-up, the small dataset cannot be used to draw conclusions. It is the authors' belief that higher acuity surgery is best followed up in clinic, both to maximize the quality of the examination and to facilitate rapid intervention toward complications should it be needed.
In conclusion, the use of telemedicine for the first postoperative visit after oculofacial plastic surgery is well-received by patients, does not appear to be associated with substantial complications, and can be augmented successfully with a web-based tool that provides high-quality photographs and visual acuity measurements. Future directions for research include building more streamlined tools for clinical application and validating the findings of this study.