Application to identify instances of anhedonia (inability to experience pleasure from activities usually found enjoyable).

Development approach: Machine-learning. Classification of past or present symptom: Both. Classes produced: Positive

The output includes- Positive mentions include:

“ X had been anhedonic”, “ X has anhedonia”.

Excludes Negative mentions, e.g

“no anhedonia”, “no evidence of anhedonia”, “not anhedonic”,

Exclude ‘Unknown’ mentions e.g:

i) used in a list, not applying to patient (e.g. typical symptoms include …); ii) uncertain (might have anhedonia, ?anhedonia, possible anhedonia); iii) not clearly present (monitor for anhedonia, anhedonia has improved); iv) listed as potential treatment side-effect; v) vague (‘she is not completely anhedonic’, ‘appears almost anhedonic’) Definitions: Search term(s): anhedon

Cohen's k=85% (50 un-annotated documents - 25 events/25 attachments, search term 'anhedon*’). Instance level, (testing done on 100 random documents):

Precision (specificity / accuracy) = 93% Recall (sensitivity / coverage) = 86% Patient level – All patients with primary diagnosis code F32 or F33 (testing done on 30 random document, one document per patient) Precision (specificity / accuracy) = 87%