Few things are read as instinctively as a human face, and few losses are felt as acutely as the loss of its movement. Severe peripheral facial palsy reshapes how a person eats, speaks, closes an eye, and is seen by others, and it carries a measurable psychological burden. Facial nerve decompression remains a valuable option in this setting—most clearly after temporal bone trauma, when electrophysiological testing points to a poor chance of spontaneous recovery^1,2—and modern middle-fossa and transmastoid techniques have made the intratemporal course of the nerve reliably accessible.³ The operation itself has a long lineage, from the first descriptions of complete decompression⁴ to the systematic rehabilitation of the traumatically injured nerve.⁵

Yet an operation is only as meaningful as our ability to measure what it achieves, and here otology has never quite settled on a single language. The House-Brackmann scale is still the most widely spoken^6,7, prized for its simplicity, but its six coarse grades flatten the very differences—resting tone, voluntary excursion, synkinesis—that matter most to a recovering patient. The Sunnybrook Facial Grading System was designed to be more sensitive, weighting these components into a single composite score^8,9, and the clinician-graded electronic facial function scale (eFACE) went further still, capturing static, dynamic, and synkinetic domains on continuous 0–100 scales.^10,11 Repeated calls for uniformity¹² and dedicated synkinesis metrics¹³ reflect how unsettled the field remains.

Into this landscape has come the smartphone. Automated, application-based grading promises to put a reproducible eFACE in the hands of any clinician—or patient—with a camera, but its trustworthiness against expert assessment is not yet established. Recent work has shown that even the manual eFACE and Sunnybrook behave differently in the hands of trained versus untrained raters¹⁴, which makes it all the more important to ask a simpler, practical question before an app enters routine use: when an algorithm grades the face, does it agree with the surgeon? We addressed this question in a cohort of patients followed after facial nerve decompression, comparing an automated eFACE against clinician-graded eFACE and the Sunnybrook.

Materials And Methods

Study design and patients

We conducted a single-center, cross-sectional study at a tertiary referral center. Sixteen consecutive patients who underwent facial nerve decompression for peripheral facial palsy and attended structured post-operative facial-function assessment were included. Patients with incomplete follow-up or without a recorded post-operative grading set were not analyzed. The study followed the principles of the Declaration of Helsinki, and institutional approval and patient consent were obtained.

Grading instruments

At the post-operative visit, each patient was graded with four instruments. The House-Brackmann scale provided a global six-grade severity anchor^6,7. The Sunnybrook composite score (0–100) was computed from resting symmetry, voluntary movement, and synkinesis^8,9. A clinician-graded (manual) eFACE was completed by an experienced examiner across its static, dynamic, and synkinetic subdomains, yielding a total score of 0–100.¹⁰ Finally, an automated eFACE total was generated by the mobile application from a standardized photographic and video set acquired at the same visit, under the same lighting and instructions used for the clinician assessment—an approach consistent with reported in-person versus video equivalence work for the eFACE.¹⁵

Statistical analysis

Continuous scores are reported as mean ± standard deviation with ranges. Relationships between instruments were quantified with Pearson and Spearman correlation coefficients. Agreement was assessed with Bland-Altman analysis—reporting the mean bias and 95% limits of agreement (bias ± 1.96 SD)—and with the two-way random-effects, absolute-agreement intraclass correlation coefficient (ICC).²² A two-sided p < 0.05 was considered significant. Analyses were performed in Python (SciPy).

Results

The cohort comprised 16 operated patients—13 with a traumatic and 3 with an idiopathic etiology—with palsy of the right side in 8 and the left in 8, and pre-operative House-Brackmann grades ranging from I to IV. Post-operative delay at assessment ranged from the immediate post-operative period to 66 months. Complete grading with all three quantitative instruments (Sunnybrook, clinician eFACE, automated eFACE) was feasible in 13 of the 16 patients (81%); the remaining three could not be graded with symmetry-based instruments because of bilateral or profoundly severe palsy and were therefore excluded from the agreement analysis. Patient characteristics and scores are summarized in Table I.

Across the 13 gradable patients, mean scores were 72.4 ± 16.3 for the Sunnybrook, 78.4 ± 9.0 for the automated eFACE, and 87.0 ± 8.3 for the clinician eFACE. The three instruments correlated strongly with one another (Fig. 1): the automated eFACE tracked the clinician eFACE with a Pearson r of 0.89 (Spearman ρ = 0.89), the clinician eFACE and the Sunnybrook agreed most closely (r = 0.93; ρ = 0.94), and the automated eFACE correlated with the Sunnybrook at r = 0.90 (all p < 0.001). Correlation was preserved across both traumatic and idiopathic patients.

Figuee 1. Correlation between grading instruments in the 13 gradable patients. (A) Automated versus clinician-graded eFACE. (B) Sunnybrook versus clinician-graded eFACE. Solid lines are least-squares regression fits; dashed grey lines are lines of identity.

Circles denote traumatic and open triangles idiopathic etiology. Pearson r is annotated; all p < 0.001. Correlation, however, is not agreement, and the Bland-Altman analysis told the more clinically important story (Fig. 2). The automated eFACE scored, on average, 8.6 points lower than the clinician eFACE (95% limits of agreement −16.7 to −0.6), a consistent, one-directional underestimation across the range of severity; the corresponding ICC was moderate (0.60). The clinician eFACE in turn ran above the Sunnybrook by 14.6 points (95% limits of agreement −3.2 to 32.5), reflecting the known scaling difference between the two instruments rather than disagreement about rank order. Representative automated eFACE outputs for two of our patients are shown in Figure 3.

Figure 2. Bland-Altman agreement plots. (A) Automated versus clinician eFACE: the automated score is systematically lower (mean bias −8.6). (B) Clinician eFACE versus Sunnybrook. Solid coloured lines mark the mean bias; dashed grey lines mark the 95% limits of agreement (bias ± 1.96 SD). Symbols as in Figure 1.

Figure 3. Representative automated eFACE application outputs for two study patients.

Table I. Facial-function scores of the 13 patients graded with all three instruments.

HB = House-Brackmann grade; mo = months; post-op = immediate post-operative assessment. Delays are approximate.