Using large language models to identify prediagnostic clinical features of ovarian cancer from healthcare records: a population-based case–control study

Garth Funston; Namu Park; Matthew Thompson; Meliha Yetisgen; Barbara A Goff; Larry Kessler; Fiona M Walter

doi:10.3399/BJGP.2025.0366

Article Figures & Data

Figures

Tables

Download figure
Open in new tab
Download powerpoint
Figure 1. Case and control participant selection. Case participants were defined by confirmed ovarian cancer diagnosis in the SEER registry. Additional inclusion criteria, including evidence of diagnostic investigation or referral in ambulatory care records, were applied to ensure availability of relevant prediagnostic clinical information for analysis. CT = computed tomography. EHR = electronic health record. ICD = International Classification of Diseases. SEER = Surveillance, Epidemiology, and End Results. US = ultrasound. UW = University of Washington.
Download figure
Open in new tab
Download powerpoint
Figure 2. Associations between coded and large language model (LLM)-extracted clinical features with ovarian cancer and early-stage (I–II) ovarian cancer. Data are odds ratios (ORs) and 95% confidence intervals (CIs). Pelvic mass OR for early stage (coded and LLM extraction not shown OR = 67.3, 95% CI = 32.0 to 141.4, see Supplementary Table S4).

Tables

Figures

View popup

Table 1. Characteristics of women with ovarian cancer (case group) and control group in ambulatory care

Characteristic	Case group(N = 136)	Control group(N = 1360)
Age, years, mean (SD)	60.7 (13.03)	60.3 (13.19)
Race/ethnicity
White	104 (76.5)	927 (68.2)
Asian	13 (9.6)	108 (7.9)
African American	8 (5.9)	95 (7.0)
Other/unknown	7 (5.1)	188 (13.8)
White Hispanic	4 (2.9)	42 (3.1)
Number of clinical visits before diagnosis/index date, median (IQR)	13.0 (8.0–25.2)	7.0 (3.0–17.0)

Data are n (%) unless otherwise specified. IQR = interquartile range. SD = standard deviation.

View popup

Table 2. Frequency of coded and large language model-extracted clinical features in case and control groups

	Case group (N = 136)n (%)			Control group (N = 1360)n (%)
Clinical feature	Identified from coded data	Identified from free text	Identified from either coded or free text	Identified from coded data	Identified from free text	Identified from either coded or free text
Abdominal
Ascites	12 (8.8)	25 (18.4)	30 (22.1)	16 (1.2)	19 (1.4)	29 (2.1)
Abdominal distention/bloating	14 (10.3)	70 (51.5)	71 (52.2)	43 (3.2)	248 (18.2)	266 (19.6)
Abdominal mass	60 (44.1)	58 (42.6)	82 (60.3)	35 (2.6)	64 (4.7)	89 (6.5)
Abdominal pain	40 (29.4)	87 (64.0)	92 (67.6)	172 (12.6)	351 (25.8)	429 (31.5)
Change in bowel habits	26 (19.1)	73 (53.7)	78 (57.4)	213 (15.7)	415 (30.5)	497 (36.5)
Early satiety	3 (2.2)	33 (24.3)	34 (25.0)	7 (0.5)	53 (3.9)	58 (4.3)
Loss of appetite	1 (0.7)	49 (36.0)	49 (36.0)	17 (1.3)	196 (14.4)	200 (14.7)
Pelvic
Pelvic mass	58 (42.6)	66 (48.5)	84 (61.8)	33 (2.4)	27 (2.0)	53 (3.9)
Pelvic pain	40 (29.4)	42 (30.9)	64 (47.1)	215 (15.8)	140 (10.3)	299 (22.0)
Vaginal bleeding	20 (14.7)	36 (26.5)	38 (27.9)	88 (6.5)	224 (16.5)	254 (18.7)
Urinary
Dysuria	12 (8.8)	20 (14.7)	27 (19.9)	147 (10.8)	196 (14.4)	286 (21.0)
Urinary frequency	19 (14.0)	44 (32.4)	51 (37.5)	111 (8.2)	200 (14.7)	248 (18.2)
Urinary incontinence	7 (5.1)	24 (17.6)	29 (21.3)	94 (6.9)	201 (14.8)	242 (17.8)
Urinary urgency	3 (2.2)	25 (18.4)	25 (18.4)	45 (3.3)	155 (11.4)	177 (13.0)
Non-specific
Fatigue	32 (23.5)	80 (58.8)	87 (64.0)	248 (18.2)	490 (36.0)	572 (42.1)
Nausea	21 (15.4)	63 (46.3)	68 (50.0)	131 (9.6)	425 (31.3)	469 (34.5)
Weight loss	5 (3.7)	47 (34.6)	47 (34.6)	53 (3.9)	283 (20.8)	303 (22.3)
Summary
Any clinical feature (≥1 of 17 symptoms)	102 (75.0)	126 (92.6)	131 (96.3)	629 (46.3)	820 (60.3)	931 (68.5)

Data are n (%) unless otherwise specified.

Supplementary Data

Supplementary material is not copyedited or typeset, and is published as supplied by the author(s). The author(s) retain(s) responsibility for its accuracy.