Prediction of individuals at high risk of chronic kidney disease during treatment with lithium for bipolar disorder

Lithium is the most effective maintenance treatment for bipolar disorder and is first-line in all international clinical practice guidelines [1]. However, its use has been declining globally [2]. Reasons for this include the required monitoring due to its narrow therapeutic window and concerns about adverse effects, particularly irreversible kidney failure [3]. In fact, kidney failure is rare [4], with end-stage kidney failure occurring at similar rates to those treated with other mood stabilisers [5], but more commonly than the general population (0.23% vs 0.11% [6]). Bipolar disorder itself appears to be associated with increased risk of kidney failure independent of lithium exposure [7]. There are inconsistencies in the existing literature about the association between kidney failure and lithium treatment duration and episodes of lithium toxicity [8].

Being able to identify individuals at risk of compromised kidney function would have high clinical utility; it would encourage the use of this effective treatment in those at low-risk and so improve outcomes for people with bipolar disorder. In the general population, established risk factors for CKD include age, sex (increased in women), ethnicity (increased in Black, Asian and Minority Ethnic (BAME) populations), family history of kidney disease, smoking, obesity, hypertension, diabetes mellitus, excessive alcohol consumption and acute kidney injury [9]. Prediction models have been developed for end-stage kidney failure in groups with a range of underlying risk [10‐14]. These tend to include a small number of core features including age, gender, ethnicity, eGFR and albuminuria. Models then vary in terms of additional features such as glucose, blood pressure, haemoglobin, lipids, calcium and phosphate. It is unclear if features related to mental health are useful in predicting CKD risk at the point of lithium initiation. It is also likely that risk factors for CKD cluster differently in patients with bipolar disorder prescribed lithium [8] so we cannot assume that risk prediction models for CKD that are of value in the general population would apply to people with bipolar disorder receiving lithium. Because CKD requiring clinical intervention (CKD stage 4 or more severe – eGFR < 30 mL/min/1.73 m²) is a rare and late-stage outcome we aimed to develop a model to classify individuals into high-risk and low-risk trajectories of kidney function following lithium treatment initiation.

We identified 1609 patients in the development sample (Aurum cohort), with a median of 14 (IQR 7–26) eGFR test results each. The median length of lithium treatment was 1.42 years (IQR 0.53–3.58). Of these patients, 401 (24.92%) developed CKD stage 3a or more severe (eGFR < 60 mL/min/1.73 m²), 38 (2.36%) CKD stage 3b (eGFR< 45 mL/min/1.73 m²), but none developed CKD stage 4. In total, 158 (9.82%) died during follow-up.

To categorise risk groups based on eGFR trajectories we chose a 5-group model, all groups with cubic trajectories (BIC = 3566.99). This defined 11.87% of the cohort as high risk. Models with 6 groups had lower BIC but included one group with less than 10% of the cohort.

Trajectories of the high-risk vs other groups (combined group 2–5) are shown in Fig. 1 and described in Table 1. Of those in the high-risk group 168 (87.96%) develop CKD stage 3a or more severe, and 25 (13.09%) developed stage 3b or more severe, compared to 233 (16.43%) and 13 (0.92%) respectively in the low-risk group.

Patients in the high-risk group were more likely to be female, younger at lithium initiation, have a lower eGFR before starting lithium and be obese. They were more likely to have a pre-existing diagnosis of migraine. They were more likely to have a record of high LDL cholesterol. They were less likely to have had an eGFR < 60 mL/min/1.73 m² any time before baseline.

There was no statistical evidence of a difference in duration of lithium treatment and incidence of lithium toxicity (> 1.5 mmol/L) between groups. Those in the high-risk group were less likely to die during follow-up and had fewer eGFR tests in total.

We used 44 features known to the clinician prior to lithium initiation to generate a prediction model for being in the high-risk group. Elastic net with 10-fold cross-validation fitted a model with λ = 0.014 and α = 1.00. The ROC area = 0.868 (95%CI 0.844–0.891) (Fig. 2). The empirical optimal cut-point was 0.134 with a sensitivity of 0.86 (95%CI 0.78–0.94) and a specificity of 0.73 (95%CI 0.63–0.84). The Youden index was 0.589. This gave a prediction accuracy of 74.54% (Table 2).

Features retained in the model were (in order of coefficient size): baseline eGFR, sex, sex by BMI group interaction, baseline eGFR by age interaction, hypothyroidism, migraine, BMI group, SSRI exposure, high LDL cholesterol, BAME, hyperthyroidism, smoking status, type 2 diabetes mellitus, and self-harm. The 3-variable KFRE and the 3-variable elastic net model performed similarly well to the full model: ROC area = 0.828 (95%CI 0.801–0.855) and ROC area = 0.852 (95%CI 0.827–0.876), respectively (Table 2).

As far as we are aware, this is the first model developed to predict high risk of future eGFR decline in people with bipolar disorder prescribed lithium. We used a large representative sample of people with bipolar disorder initiated on lithium and followed up for up for a median of 7.10 years (IQR 3.85–11.36). It is also the first study to use the two CPRD datasets, covering a large, representative sample of the UK population to develop a prediction model and provide external validation.

Because of the rarity of kidney failure, and the varied follow-up time and eGFR recording frequency in EHRs, we sought to identify approximately 10% of individuals prescribed lithium who were at highest risk of deteriorating kidney function, defined by the trajectory of their serial eGFR measurements. Using group-based trajectory analysis we identified high-risk groups independently in the Aurum and Gold cohorts. In both cases approximately 85% of those categorised as high risk developed CKD stage 3a or more severe compared to approximately 15% in the low-risk groups. Approximately 10% of those identified as high risk developed CKD stage 3b or more severe, compared to < 1% in the low-risk group. A number of features differed between the high-risk and low-risk groups in both cohorts. Those in the high-risk groups were more likely to be female, younger, more likely to have a lower eGFR before starting lithium, more likely to experience migraine and less likely to have a prior record of eGFR < 60 mL/min/1.73 m². These CKD risk factors have been previously identified. CKD is more common in women [25], and this has been shown in lithium users specifically [26]. Younger women appear to be at particular risk [26]. Low baseline eGFR increases risk of CKD in the general population [27]. Migraine is not commonly thought of as a risk factor for CKD, but has been identified as such in one study, especially in younger age groups [28]. Migraine is highly comorbid with bipolar disorder [29] and it may also be a proxy for medication use which could impair kidney function. In both cohorts, there was no association between duration of lithium treatment or lithium toxicity (which was rare) and high-risk group membership.

Our model, developed in CPRD Aurum to predict whether individuals were in a high-risk group for eGFR decline during treatment with lithium for bipolar disorder, had excellent discrimination in the CPRD Gold cohort (ROC area = 0.879). However, simple models only including sex, age and baseline eGFR performed similarly well (3-variable KFRE ROC area = 0.870, 3-variable elastic net ROC area = 0.888), all with similar levels of accuracy (> 75%). In the external validation data set, our model designed to predict high-risk trajectory predicted CKD stage 3b or more severe (eGFR < 45 mL/min/1.73 m²) with a ROC area = 0.849. Again, simple models performed well, with the 3-variable KFRE having the highest accuracy (78%).

When we reduced our cohort to those starting lithium with an eGFR ≥ 90 mL/min/1.73 m² our full model and 3-variable elastic net model performed better than the KFRE. However, our sample was too small to complete external validation of these new models.

Given these findings, simple risk calculators should be used in clinical practice at the decision to commence lithium and when eGFR is regularly measured. This could be the KFRE or our 3-varaible elastic net model (Eq. 1), which performs better than the 3-variable KFRE when eGFR ≥ 90 mL/min/1.73 m². We did not find predictors of eGFR decline that were specific to lithium-treated patients.

We developed a model for predicting, at lithium initiation, individuals at high risk of a poor trajectory of kidney function using serial eGFR measurements. We externally validated this model, which had excellent discrimination and good calibration. CKD stage 3b or more severe occurred in 2% of the population across the two cohorts. Whilst this is worrying, it means that the vast majority of patients treated with lithium do not develop kidney failure, and those at risk can be identified prior to initiating lithium using their age, sex and baseline eGFR.