Characterising the transdiagnostic prodrome to severe mental disorders with natural language processing: an electronic health record study: Abstracts of the 36th ECNP Congress 2023

Background: Preventive approaches for severe mental disorders (SMD) rely on accurate knowledge of the duration and first presentation, timecourse and transdiagnosticity of prodromal features. The prodromal phase has mostly been investigated in “look-back” studies which have several limitations including the use of interviews that are not integrated into clinical practice, small sample sizes and substantial recall biases. Moreover, it is unclear to what extent SMD may display diagnostic-specific or transdiagnostic prodromal symptoms. To address these limitations, we will take advantage of Electronic Health Records (EHRs) routinely employed in secondary mental healthcare, constituting an extensive pool of untapped real-world information. We aim to describe and compare the (1) incidence and duration of the first prodromal clusters, (2) mean number of occurrences of prodromal clusters between SMD diagnostic groups and prodromal years, and (3) total number of occurrences of prodromal features.

Methods: Clinical register-based EHR cohort including all individuals accessing the South London and the Maudsley (SLaM) services (2008-2021) and receiving a primary (i.e. not comorbid) diagnosis of any ICD-10 mental disorder that falls under SMD. These diagnoses were then clustered into diagnostic groups: non-bipolar mood disorders (NBMD), bipolar disorders (BD), and psychotic disorders (PSY). We extracted 52 prodromal features (symptom and substance use variables) through NLP algorithms that are pooled in 8 broader prodromal clusters (depressive, disorganised, catatonic, manic, negative, positive, substance use and other symptoms).

Prodromal features and clusters were extracted monthly from the index date of diagnosis until their first occurrence. For aim 1, the duration (months) and presentation (incidence) of the first prodromal clusters were compared between SMD groups with Cohen’s F (f) and D (d) effect sizes derived from one-way ANOVA. For aim 2, the timecourse (annualised mean occurrence) of prodromal clusters was compared between SMD groups over 12 prodromal years using a linear mixed-effects model. For aim 3, the mean number of occurrences of each prodromal feature in the prodromal period was compared with three-wise and pair-wise discriminability scores between SMD groups to assess symptom transdiagnosticity.

Results: 26,975 individuals were included (NBMD=13,422; BD=2,506; PSY=11,047; mean age 41.8 years [SD=17.4]; 55% female; 52% white self-assigned ethnicity). The duration of prodromal period was shorter in NBMD (median[IQR]= 18[36] months) than BD (26[35], d=0.21) and PSY (24[38], d=0.18). The most common first prodromal cluster was other symptoms (88% NBMD, 85% BD, 78% PSY); there were negligible/small between-group differences in the incidence of all eight first prodromal clusters (0.01<f<0.32). The annualised mean occurrence of depressive and other symptoms (across all SMD groups), and positive symptoms (in PSY), increased approaching the SMD onset. All symptoms showed negligible/small discriminability across all three SMD groups, with paranoia showing the highest discriminability (f=0.37). Several prodromal features showed medium to large discriminability (0.50<d<0.72) across paired SMD groups.

Discussion: This large AI-NLP-based analysis identified scarce differences in the duration and first presentation of the prodrome across NBMD, BD and PSY. Depressive and other symptoms intensified when approaching SMD onset and most prodromal features display transdiagnosticity across SMD groups. These findings inform early detection and prevention.

Conflict of interest:

Disclosure statement: MA is supported by the UK Medical Research Council (MR/N013700/1) and King’s College London member of the MRC Doctoral Training Partnership in Biomedical Sciences.