Abstract

Background: Variation in speech may be an early sign of minimal hepatic encephalopathy (MHE) or future overt HE (OHE). Though a battery of tests is available for MHE assessment, these are cumbersome and rarely used in clinical practice. In a prospective study, we evaluated the ability of speech recorded at home or in the office to correlate with validated HE assessments in patients of diverse backgrounds.

Methods: In a prospective study (“HE Audio Recording to Detect MHE” or HEAR-MHE), we enrolled 212 patients (169 cirrhosis and 43 non-cirrhosis controls) from two geographically disparate centers. Patients underwent psychometric HE score (PHES; validated test to diagnose MHE), animal naming test, and audio recording while reading a paragraph. Speech variables (acoustic, lexical, and syntactic) were automatically extracted from recordings via the Winterlight Labs Smartphone analysis platform, an app initially designed to characterize speech in dementia, and immediately transmitted data to the research team via a secure server. Patients with cirrhosis were in 3 non-overlapping categories: (1) prior OHE + HE treatment, (2) MHE with no prior OHE (diagnosed by PHES ≤ -4), (3) no MHE or prior OHE. T-tests compared continuous variables, Pearson correlation correlated continuous variables, and Vuong’s test compared AUCs of different models.

Results: Cirrhosis: median 63 years (IQR 55, 68), 52% male, median MELD 9 (IQR 7, 12), 39% alcohol and 33% fatty liver. Controls: median 55 years (IQR 42, 62), 56% male, and 58% had fatty liver disease and 16% viral hepatitis.

Office Audio: Audio recordings were median 36 seconds (IQR 33, 41). Speech rate was significantly slower in patients with MHE (152 words/min) and history of OHE (155 words/min) as compared to either no MHE or controls (169 and 172 words/min; Figure). Prediction of MHE using speech rate was comparable to the animal naming test (AUC 0.73 vs 0.66, P=0.12)

Home Audio: Patients performed 43 audio recordings with personal Smartphone apps. 23/225 (10%) speech variables were significantly correlated (<0.05) between home and office recordings, despite recordings on different days. Speech rate at home was highly correlated to that in the office (r=0.53, P<0.01) and PHES score in office (r=0.71, P<0.01).

Speech rate findings were similar regardless of site (p=0.44) or patient region of origin (p=0.30), in the whole cohort or within patient sub-groups.

Conclusion: Brief audio recordings through a Smartphone app either at home or office may help identify patients with MHE and perform as well as traditional HE tests. Speech data can be reliably obtained at home or in the office and is not adversely impacted by accent. Automated speech analysis through Smartphone apps is a promising novel method to identify high risk patients at risk for future complications of decompensated cirrhosis.