Abstract CE-50 Table 1

Performance of traditional structured definitions and supervised machine learning algorithms in case identification of SLE in the electronic health record

Structured Definitions				Supervised Machine Learning

	Recall		Precision		Recall		Precision

	(Sensitivity)	Specificity	(PPV)		(Sensitivity)	Specificity	(PPV)
Single ICD9 710.0	0.99	0.97	0.79	All ICD-9 codes and counts¹	0.89	0.99	0.89
Single ICD9 710.0 + any lupus medication	0.96	0.98	0.86	All ICD-9 codes and counts + NLP of clinical notes²	0.90	0.99	0.89
Single ICD9 710.0 + any lupus medication + any positive lupus-serology	0.93	0.98	0.87	All ICD-9 codes and counts + NLP of clinical notes + all serologic data³ + all medication data	0.91	0.99	0.92
				All ICD-9 codes and counts + NLP of clinical notes + all serologic data + all medication data + demographics⁴	0.85	0.99	0.96

↵1 Supervised Machine Learning algorithms included all available ICD-9 codes for patients as well as counts and locations in the medical records in which they were found (i.e. clinical encounters, problems lists, medications orders, etc.)
↵2 All text data from clinical notes associated with a patient’s medical record were included in the ML algorithm
↵3 Serologic data included ANA, double-stranded DNA, anti-Smith antibody, anti-RNP, SSA, and SSB
↵4 Demographic information included age, gender, race/ethnicity, insurance status, and employment status