Fralick Michael, Højbjerg Lassen Mats C, Rangrej Jagadish, Asgari Sahar, Rais Saad, Hillmer Michael P, Fritz Jamie Lee, Zorcic Katarina, Perkins Bruce A, Colacci Michael, Biering-Sørensen Tor, Mamdani Muhammad, Campbell Kieran R
Sinai Health System, Division of General Internal Medicine, Toronto, Ontario, Canada; Institute for Health Policy, Management, and Evaluation, University of Toronto, Toronto, ON, Canada; Lunenfeld-Tanenbaum Research Institute, Sinai Health System, Toronto, Ontario, Canada.
Department of Cardiology, Copenhagen University Hospital - Herlev & Gentofte, Copenhagen, Denmark.
J Diabetes Complications. 2025 Oct;39(10):109144. doi: 10.1016/j.jdiacomp.2025.109144. Epub 2025 Jul 30.
Sodium glucose co-transporter 2 inhibitors (SGLT2i) are associated with a small-magnitude but higher risk of diabetic ketoacidosis (DKA). However, objectively identifying patients at lowest and highest risk of DKA is challenging.
We developed a prediction model using outpatient prescription data from Ontario, Canada and externally validated it using data from Denmark. We included adults with type 2 diabetes mellitus who were newly prescribed an SGLT2i. Our candidate predictors in the model were based on prior work and included the following: Sex, insulin use, prior DKA, dementia, hemoglobin A1C, and creatinine. Our outcome was 1-year risk of hospitalization with DKA. We calculated a risk score using an adaptation of penalized regression for each patient reported test characteristics in Ontario (derivation cohort) and Denmark (external validation cohort).
We identified 322,135 in Ontario and 43,377 adults in Denmark who had type 2 diabetes mellitus and received an SGLT2i. The absolute risk of DKA within 1-year was 0.28 % (N = 916) in Ontario and 0.23 % (N = 101) in Denmark. Using data from Ontario, the risk score for each variable were as follows: Insulin use = 4 points, A1C > 9 % = 4 points and prior DKA = 19 points. All other variables received zero points. The overall model AUC was 63 % in Ontario and 66 % in Denmark (external validation set). Within Ontario, at a score threshold of zero, the risk of DKA was 0.19 % and the PPV was 0.3 % and the sensitivity was 100 % and similar results were observed in Denmark. For adults with a score of 19 or higher, the risk of DKA was 35-fold higher but false positives were common yielding a PPV of 6.7 % and sensitivity was lower at 3 %. In Denmark, adults with a score of 19 or higher had a risk of 11 % and the PPV was 10.2 % and sensitivity was 5 %.
Adults with a score of 0 (that is, simply a lack of DKA history, lack of insulin therapy, and A1c < 9 %) can be reassured that 99.8 % will not experience DKA in the subsequent year. In contrast, for adults with a score of 19 or higher the one-year risk of DKA is approximately 9 %, but false positives and false negatives are common and thus more work is needed to improve the predictive performance of the model.
钠-葡萄糖协同转运蛋白2抑制剂（SGLT2i）与糖尿病酮症酸中毒（DKA）风险小幅升高相关。然而，客观识别DKA风险最低和最高的患者具有挑战性。
我们利用加拿大安大略省的门诊处方数据开发了一个预测模型，并使用丹麦的数据进行外部验证。我们纳入了新开具SGLT2i处方的2型糖尿病成年患者。模型中的候选预测因素基于先前的研究工作，包括以下因素：性别、胰岛素使用情况、既往DKA病史、痴呆、糖化血红蛋白A1C和肌酐。我们的研究结果是DKA住院的1年风险。我们使用一种惩罚回归的变体为安大略省（推导队列）和丹麦（外部验证队列）中报告了测试特征的每位患者计算风险评分。
我们在安大略省识别出322,135例、在丹麦识别出43,377例患有2型糖尿病并接受SGLT2i治疗的成年人。安大略省1年内DKA的绝对风险为0.28%（n = 916），丹麦为0.23%（n = 101）。利用安大略省的数据，每个变量的风险评分如下：胰岛素使用 = 4分，A1C > 9% = 4分，既往DKA病史 = 19分。所有其他变量得0分。模型在安大略省的总体曲线下面积（AUC）为63%，在丹麦（外部验证集）为66%。在安大略省，风险评分为零的患者中，DKA风险为0.19%，阳性预测值（PPV）为0.3%，敏感性为100%，丹麦也观察到类似结果。对于风险评分为19分或更高的成年人，DKA风险高出35倍，但假阳性很常见，PPV为6.7%，敏感性较低，为3%。在丹麦，风险评分为19分或更高的成年人风险为11%，PPV为10.2%，敏感性为5%。
风险评分为0分（即仅缺乏DKA病史、未接受胰岛素治疗且A1c < 9%）的成年人可以放心，99.8%的人在随后一年不会发生DKA。相比之下，对于风险评分为19分或更高的成年人，DKA的1年风险约为9%，但假阳性和假阴性很常见，因此需要做更多工作来提高模型的预测性能。
