Steeper on current exposure than on cumulative: valproate and PCOS in a Danish register
- Nationwide register cohort study from Denmark, run from the psychiatry department of Aalborg University Hospital with co-authors in Aarhus and Hjoerring, using the Danish Psychiatric Central Research Register, the National Patient Register and the National Prescription Registry. All women in Denmark with a first diagnosis of bipolar disorder (ICD-10 F30.x to F31.x) or epilepsy (ICD-10 G40.x) between 1 January 2000 and 31 July 2022 were included; women with either diagnosis, with valproate exposure or with polycystic ovary syndrome before 1 January 2000 were excluded. N = 20,967, of whom 8,110 had bipolar disorder, 13,006 had epilepsy and 149 had both. Total follow-up was 211,224 person-years, 55,275 of them in the bipolar group and 155,950 in the epilepsy group. Median age at bipolar diagnosis was 26.5 years, median age at a PCOS diagnosis in that group 28.7 years. The outcome was an incident coded diagnosis of polycystic ovary syndrome (ICD-10 E28.2) in the national patient register. Analyses were run in STATA 18 with a significance threshold of 0.05.
- The analysis the Methods declare as primary is current cumulative valproate exposure in the pooled bipolar and epilepsy cohort. Current cumulative exposure runs on a 90-day rolling window opened by every dispensed prescription: defined daily doses inside the window are summed, and the count resets to zero once more than 90 days pass without a new dispensing. In that pooled cohort, hazard rate ratios against no current exposure were 4.43 (95% CI 3.42 to 5.73), 6.86 (95% CI 4.19 to 11.23) and 7.08 (95% CI 3.85 to 13.03) across rising dose bands, all p < 0.001.
- The bipolar figures are a stratified, that is secondary, analysis and appear only in Table 2, which carries the current cumulative model alone: 4.66 (95% CI 2.99 to 7.25) for up to 90 defined daily doses, 5.89 (95% CI 2.74 to 12.65) for 90 to 365, and 8.76 (95% CI 3.11 to 24.69) above 365, all p < 0.001. In the bipolar group 91 women received a PCOS code and 51 of them had been exposed to valproate. Event counts by dose band are not published; our reading that the top band stands on few events is inferred from the width of its interval.
- Counted the other way the association is smaller but does not disappear, and it belongs to a different population: overall cumulative exposure, the total defined daily doses dispensed since cohort entry, is published for the pooled bipolar and epilepsy cohort only. It gave 1.35 (95% CI 1.03 to 1.77, p < 0.05), then 1.25 (95% CI 0.76 to 2.05, p = 0.376, not significant), then 2.04 (95% CI 1.29 to 3.22, p < 0.01), which is significant. The ever versus never comparison, also pooled, gave 1.55 (95% CI 1.20 to 2.00, p = 0.001).
- The authors state their own conclusion as follows: valproate exposure was associated with an increased risk of polycystic ovary syndrome in women with bipolar disorder or epilepsy, with the strongest associations observed for current cumulative exposure and additional risk apparent at high cumulative dosages; the Discussion adds that overall cumulative exposure showed a less clear dose-response pattern. Their closing paragraph reinforces existing guideline advice to avoid valproate in females of reproductive potential whenever possible, and to arrange reproductive and metabolic monitoring when it is used.
- Nobody rated the outcome and nobody was blinded. Exposure and outcome both come from administrative registers: valproate from redeemed prescriptions, PCOS from codes entered by treating clinicians in routine care. The Strengths section calls the registers themselves validated; what the publication does not describe is any independent adjudication of the outcome or any validation specific to the PCOS code.
Maintenance treatment in bipolar disorder is normally judged by an outcome the patient never feels directly: an episode that did not happen. This study counted the other side of the ledger: an endocrine diagnosis that arrives while the patient is stable, counted in a way that dates its arrival.
Two ways of counting, two gradients
Valproate exposure was modelled three ways, and the three counts do not agree; the Discussion is built on that disagreement. Current cumulative exposure opens a 90-day window at every redeemed prescription, adds up the defined daily doses inside it, and resets to zero when more than 90 days pass without a new redemption, so it describes whether a woman is on the drug now and roughly how much. Overall cumulative exposure adds every defined daily dose dispensed since cohort entry and never resets, so it describes a treatment career.
In the bipolar stratum, current cumulative exposure gives 4.66 (95% CI 2.99 to 7.25), 5.89 (95% CI 2.74 to 12.65) and 8.76 (95% CI 3.11 to 24.69) across rising bands. Overall cumulative exposure, published for the pooled bipolar and epilepsy cohort only, gives 1.35 (95% CI 1.03 to 1.77), then 1.25 (95% CI 0.76 to 2.05, p = 0.376), then 2.04 (95% CI 1.29 to 3.22, p < 0.01). The top band there is significant, so cumulative exposure does not drop out of the picture; what differs is the shape. On the current count the bands rise in order; on the cumulative count the middle band falls below the lowest and misses significance. The authors offer two readings: risk may be driven by acute pharmacological effects that fade once treatment stops, or a cumulative threshold may have to be crossed before a diagnosis is written down. Both are theirs, and both are put as possibilities.
The bipolar numbers here are the secondary ones
The analysis declared primary in the Methods is the pooled one, and epilepsy supplies 155,950 of its 211,224 person-years. The bipolar values sit in Table 2, which carries the current cumulative model alone, for bipolar disorder and for epilepsy alike; the ever versus never and overall cumulative models are given for the pooled cohort. Setting 4.66 beside 1.35 crosses populations as well as counting methods.
Table 2 also shows that the ordered rise does not hold in every stratum. In epilepsy the bands run 4.37 (95% CI 3.18 to 6.01), 7.08 (95% CI 3.68 to 13.65) and 6.25 (95% CI 2.94 to 13.30), so the top band sits below the middle one. Inside the bipolar stratum, 91 women received a PCOS code and 51 of them had been exposed. The interval around 4.66 is the narrowest of the three bands, and it is the band a woman occupies during the first three months of treatment.
What the design cannot settle
The study is observational, and the publication gives no trial registration and no protocol reference. The authors state that residual confounding by indication and channeling bias cannot be excluded, since women prescribed valproate in their childbearing years may be a group with more severe or treatment-resistant illness. Body mass index, lifestyle and family history of PCOS were lacking. Data on co-medications were available, but adjustment for time-varying polypharmacy and prior treatment sequences was not undertaken, which the authors explain as avoiding overadjustment. The outcome is a code rather than a hormone panel.
What changes in the room
No prescribing rule follows from this. What it supplies is timing: on the count that tracks current treatment, the association is already there in the lowest band, which a woman reaches within the first three months rather than after years. No external funding was sought, and data access and analyses were covered by the ordinary budget of the psychiatric research unit at Aalborg University Hospital. Competing interests are declared in a separate section: three authors report none, one reports advisory board and lecture fees from UCB Nordic and Eisai, and the senior author reports investigator work for Compass Pharmaceuticals and Boehringer-Ingelheim and speaking fees from Lundbeck, Teva, Janssen-Cilag and Boehringer-Ingelheim.
The authors state their own conclusion this way: valproate exposure was associated with an increased risk of polycystic ovary syndrome in women with bipolar disorder or epilepsy, with the strongest associations observed for current cumulative exposure and additional risk apparent at high cumulative dosages, while overall cumulative exposure showed a less clear dose-response pattern.
This is observational register data, and the publication gives no trial registration and no reference to a protocol, so association cannot be read as causation. The authors state that residual confounding by indication and channeling bias cannot be excluded. Body mass index, lifestyle and family history were lacking, while data on co-medications were available but adjustment for time-varying polypharmacy and prior treatment sequences was not undertaken. The outcome is a diagnostic code entered in routine care rather than a hormone measurement; the Strengths section calls the registers themselves validated, and what the publication does not describe is any independent adjudication of the outcome or any validation specific to the PCOS code. The analysis declared primary pools bipolar disorder with epilepsy, which supplies 155,950 of the 211,224 person-years, and the overall cumulative model is published for the pooled cohort only, so the bipolar and the pooled figures are not one population. Event counts by dose band are not published, and our reading that the top bipolar band stands on few events is inferred from its interval, 3.11 to 24.69. The abstract splits the cohort as 8,003 and 12,964 while the Results give 8,110 and 13,006 with 149 women holding both diagnoses; the Results figures are used here, because the two reconcile if the 149 are assigned by first diagnosis, which the publication does not state. The Introduction says the study employs multiple sensitivity analyses; none is reported in the Methods, the Results, Tables 1 and 2 or the Discussion, and on the Springer publication page no supplementary materials are indicated. The abstract prints the interval 1.28 to 3.20 where Table 1 prints 1.29 to 3.22, and the table was followed here.