Two outcomes and nothing else: what Japan's national claims database counts as a result in bipolar disorder
- Population-based cohort study on the National Database of Health Insurance Claims and Specific Health Check-ups of Japan, a Ministry of Health, Labour and Welfare holding established in fiscal 2009 that covers the whole insured population, approximately 126 million people. Adults aged 20 and over with a primary diagnosis of bipolar disorder, ICD-10 codes F30 to F31, treated in psychiatric settings between 1 April 2013 and 31 March 2022, followed to 31 May 2023. Inclusion further required at least two lithium prescriptions at 200 mg per day or above, a rule that removed 501,536 of the 861,016 eligible adults, 58.2%, so the cohort is patients who have at some point received lithium rather than everyone carrying the diagnosis. The analysis set is 315,046 patients, median follow-up 7.1 years, interquartile range 3.7 to 9.7. The design is within-individual: each patient serves as their own control, estimated by stratified Cox regression adjusted for time-varying covariates. 83,621 patients (26.5%) were admitted; 208,896 psychiatric admissions were recorded, 10.2 per 100 person-years.
- The publication measures two outcomes and no others. The primary outcome is time to psychiatric hospitalisation; the secondary is time to all-cause hospitalisation, which the authors declare as a proxy measure of the overall effectiveness of treatments reflecting tolerability. The same pair runs through eight of the nine supplementary tables, psychiatric hospitalisation in 2, 4, 6, 7, 8 and 9 and all-cause hospitalisation in 3 and 5, while Supplementary Table 1 is a list of generic drug names and holds no outcome at all. Weight, cognitive function, renal and thyroid function, work capacity and quality of life are not reported in this publication. Thyroid disorder sits in the data as a baseline characteristic of the cohort, 31,825 patients (10.1%) in Table 1, and not as an outcome.
- Monotherapy against periods of non-use of the same class, adjusted hazard ratios: lithium 0.67 (95% CI 0.66 to 0.68), valproate 0.71 (0.70 to 0.73), lamotrigine 0.72 (0.69 to 0.75), carbamazepine 0.74 (0.70 to 0.78). The authors report the ordering and add that the difference against the other mood stabilisers was not large, aHR 0.67 against 0.71, 0.72 and 0.74. Their stated head conclusion is that mono- and combination therapy with mood stabilisers and antipsychotics exhibited substantial heterogeneity in their effectiveness for preventing psychiatric hospitalisation. Separately, in the discussion, they write that the order of risk reduction suggests lithium had the greatest preventive effectiveness.
- The abstract carries only the combinations that lowered risk. Against lithium monotherapy, adding carbamazepine gives 0.73 (0.64 to 0.83), zotepine 0.82 (0.72 to 0.93), aripiprazole 0.87 (0.82 to 0.92) and valproate 0.92 (0.87 to 0.97), and all four are in the abstract. Adding sulpiride to lithium gives 1.16 (1.03 to 1.30), and adding chlorpromazine to lamotrigine gives 1.47 (1.19 to 1.81). Both are significant increases, both appear in Figure 3 and in the Results text, and neither appears in the abstract. I followed Figure 3 and Supplementary Table 4.
What a claims database was built to see
The National Database of Health Insurance Claims and Specific Health Check-ups of Japan is not a research cohort. Established in fiscal 2009 under universal insurance, it holds the claims facilities submit to insurers and publicly funded programmes: diagnoses, procedures, prescriptions. Its unit of record is a payment event, which fixes what any study built on it can take as an outcome. Under Japan's reimbursement rules and the Medical Care Act, patients with mental illness are, in the paper's wording, typically admitted to designated psychiatric beds, except where admission is required for a physical comorbidity. That makes psychiatric hospitalisation a bounded administrative category, and the supplementary methods name where the boundary leaks: some patients are admitted to general hospital wards for suicide attempts without being hospitalised in psychiatric facilities. Yasuyuki Okumura at the Initiative for Clinical Epidemiological Research in Tokyo, with Hidetaka Tamune, Hiroyuki Harada and Tadafumi Kato at Juntendo University, took that category as the primary outcome for 315,046 adults, median follow-up 7.1 years.
Two outcomes, eight tables
The secondary outcome is all-cause hospitalisation, and the supplementary methods give it two rationales: a proxy measure of the overall effectiveness of treatments reflecting tolerability, and those suicide-attempt admissions to general wards that the primary outcome does not see. That is what tolerability means inside this accounting: an adverse effect registers only if it puts the patient in a bed. The authors name the gap in one sentence in the discussion, cautioning about additional adverse events, such as extrapyramidal symptoms, akathisia, weight gain and prolactin elevation, that may not require hospitalisation but could impair treatment adherence and quality of life. That sentence appears in no table. Weight, cognitive function, renal and thyroid function, work capacity and quality of life are not reported in this publication.
Thyroid disorder is the sharpest illustration, because the variable is present: Table 1 records it in 31,825 patients, 10.1% of the cohort, as a baseline characteristic describing who entered. It is not among the things the study then asks about them. The same asymmetry runs through Supplementary Table 2, which carries rows that never reach the main text: pimozide at 2.89 (1.19 to 7.00) on 272 users, a significant increase the article never names, and benzodiazepines 1.22 (1.20 to 1.25) and antidepressants 1.06 (1.04 to 1.08), also increases. In the Supplementary Discussion the last two serve as a construct-validity check, reproducing known register findings. A calibration instrument in the paper's logic, a prescription in the patient's week.
What the drugs did on the primary outcome
On that outcome the estimates are as follows. Lithium monotherapy carries an adjusted hazard ratio of 0.67 (95% CI 0.66 to 0.68), valproate 0.71 (0.70 to 0.73), lamotrigine 0.72 (0.69 to 0.75) and carbamazepine 0.74 (0.70 to 0.78), each against periods of non-use, and the authors note the spread between them is not wide. Among antipsychotics the discussion names sultopride as the greatest risk reduction at 0.53 (0.43 to 0.66), which holds inside the subset displayed in Figure 2, restricted to drugs with 1,000 or more users. Supplementary Table 2 is not so restricted and contains a lower value, clozapine 0.44 (0.37 to 0.52) on 306 users. This is a boundary of the figure rather than a fault in the arithmetic: the same publication answers which agent most reduces admissions differently depending on which table is open.
Who is inside the 315,046
Of 861,016 adults in the database with a primary bipolar diagnosis, 315,046 reached the analysis, 36.6%. The largest filter is the rule requiring at least two lithium prescriptions at 200 mg per day or above: it removed 501,536 people, 58.2%. The choice is defended: lithium in Japan is licensed only for bipolar disorder while other agents carry additional indications, so the rule raises diagnostic specificity. The Supplementary Discussion concedes the cost: some patients with genuine bipolar disorder were excluded for want of a documented lithium prescription. In the patient selection section the paper reports that 47% of Japanese patients with bipolar disorder receive lithium against 16% in the United Kingdom and 17% in the United States, but cites those figures to prior publications, references 16 to 18, not measurements of its own. This study is therefore evidence about patients who have at some point been given lithium, counted by the one event their insurer records.
The authors' head conclusion is that mono- and combination therapy exhibited substantial heterogeneity in their effectiveness for preventing psychiatric hospitalisation, and that hospitalisation, together with hospitalisation of any cause, is the whole of what this publication measures.
The design is observational and cannot establish causation. Within-individual comparison includes only patients who had both exposed and control periods and who had an outcome event, which the authors list as a limit on transportability, and the inclusion rule requiring at least two lithium prescriptions removed 501,536 of 861,016 eligible adults, 58.2%, so the findings describe patients who have at some point received lithium rather than every patient carrying the diagnosis. There is no blinding by construction: the outcome is extracted from an administrative record and no one assesses it. Residual confounding by indication remains, because medication is often started during deterioration, and the authors say so. Exposure is a prescription record, not verified intake; the paper notes that therapeutic drug monitoring would be needed to establish adherence. The reimbursement system does not record the reason for admission, so prevention of a manic episode cannot be separated from prevention of a depressive one. The authors state that the design does not adequately account for time-varying covariates such as symptom severity, psychiatric comorbidities or changes in concomitant medications other than psychotropics. Ethnicity data were unavailable. The abstract names 15 antipsychotics associated with lower risk while the Results text lists 17; the two extra, blonanserin 0.91 (0.83 to 1.00) and chlorpromazine 0.95 (0.90 to 1.00), have an upper bound of exactly 1.00 in Figure 2, and the Results text attributes the 1,000-user restriction to Figure 1 while the caption to Figure 2 carries it. Neither the main text nor the supplementary material states a protocol registration number; ethical approval is Juntendo University Ethics Committee no. E22-0372-M01, use of the database was approved by the Ministry of Health, Labour and Welfare, and the analytic code, with the code book, is open at OSF PMBCT, while the data set itself is proprietary to the Ministry, cannot be shared with third parties and is reachable only by separate application. Outcome ascertainment rests on code lists compiled by the investigators, and the supplementary outcome table records the measurement characteristics and validation of the outcome definition as n/a. Funding was the Japan Agency for Medical Research and Development, grant JP23dk0307124, declared with no role in design, collection, analysis, interpretation or writing. Competing interests are declared separately and are extensive: the senior author reports grants and personal fees from a long list of manufacturers, including Otsuka, Sumitomo Pharma and Janssen, outside the submitted work, and the first author is president of the Initiative for Clinical Epidemiological Research and a former employee of Real World Data Co. The paper does not connect these declarations to its results, and the connection is worth stating: aripiprazole, aripiprazole long-acting and brexpiprazole are Otsuka products and paliperidone long-acting is a Janssen one, all four sit among the lowest values in Figure 2, and aripiprazole is the only antipsychotic associated with lower risk when added to each of lithium, valproate and lamotrigine.