Everyone's mood varies. Bipolar disorder is not that.
It is a mood disorder defined by episodes — discrete periods, lasting days to months, in which mood, energy and activity shift away from a person's normal state and stay shifted. The defining episode is one of abnormally elevated or irritable mood with increased energy and activity: severe enough to disrupt functioning, require hospitalisation or involve psychosis, and it is called mania; milder, shorter, clearly abnormal but not disabling in the same way, and it is called hypomania. Most people with the condition also have episodes of depression. Between episodes, mood is frequently normal — a period clinicians call euthymia.
That episodic structure is the whole diagnosis. Bipolar I requires at least one manic episode, with or without depression. Bipolar II requires at least one hypomanic episode plus at least one major depressive episode, and no full mania. The distinction is not severity of the depression — bipolar II depression is if anything the more burdensome — but whether the high ever crossed into full mania.
It is not common, and it is disabling out of proportion to its frequency. Lifetime prevalence in the United States is 1.0 percent for bipolar I, 1.1 percent for bipolar II and 2.4 percent for subthreshold bipolar disorder [1]; across eleven countries in the World Mental Health surveys the aggregate figures were 0.6, 0.4 and 1.4 percent, totalling 2.4 percent for the whole spectrum [2]. Yet it accounted for 9.9 million disability-adjusted life-years in 2013 — 0.4 percent of all DALYs and 1.3 percent of all years lived with disability — with the burden appearing from age 10, peaking in the twenties, and remarkably constant across geography [3]. Early onset, chronicity and severity, not prevalence, are what make it costly.
And it is dangerous. Suicide risk is substantially elevated. A systematic review across 13 suicide studies and 23 attempt studies identified previous attempt and hopelessness as the main risk factors for suicide, with family history of suicide, early onset, depressive symptom burden, mixed states, rapid cycling, comorbid disorders and substance misuse associated with non-fatal suicidal behaviour [4] [5] [6] [7]. This review will show that one drug is associated with lowering that risk.
The clinical problem this review is organised around. People seek help when they are suffering, and mania rarely feels like suffering at the time — energy, confidence and reduced need for sleep are not symptoms most people take to a doctor. Depression is. So bipolar disorder is very frequently diagnosed and treated as unipolar depression, sometimes for years, and the treatment that follows from that mistake is different from — and sometimes worse than — the treatment the illness needs.
The evidence for how large that problem is, and the reason it happens, is quantitative and follows below.
A note on framing. Nothing in this review supports the popular caricature of bipolar disorder as dramatic, unpredictable or defining of a person. The longitudinal data say the opposite: the illness is mostly depressive, mostly subsyndromal, and when patients are asymptomatic their psychosocial functioning is good — not quite as good as controls, but good [8]. It is an episodic, treatable medical condition with long well periods, and it is described that way here.
Centerpiece: what maintenance treatment buys
Bipolar disorder is recurrent. Treating the episode in front of you is necessary and insufficient; the illness will produce another one. So the question that decides long-term outcome is not "does this drug treat mania" but "does staying on it lower the rate at which episodes come back".
The cleanest maintenance dataset is a pair of prospectively combined 18-month trials. 1,315 bipolar I patients entered an open-label phase; 638 who stabilised were randomised to lamotrigine (n=280), lithium (n=167, serum 0.8–1.1 mEq/L) or placebo (n=191), with the primary endpoint being time from randomisation to intervention for a mood episode [9].
The median times were 86 days on placebo, 184 on lithium and 197 on lamotrigine [9].
The first check the arithmetic was never given. The paper prints three medians and never divides them. Lithium multiplies the median time to the next mood episode by 2.14 and lamotrigine by 2.29. Maintenance treatment roughly doubles the interval between episodes — which is the single most useful sentence anyone can be told about why to keep taking it. In concrete terms: the placebo median was under three months, and on either active drug it was over six [9].
The second check is the more interesting one, and it explains why there is more than one mood stabiliser. The recently-depressed trial reports one-year intervention-free proportions separately by pole: for depression, lamotrigine 57 percent, lithium 46 percent, placebo 45 percent; for mania, lamotrigine 77 percent, lithium 86 percent, placebo 72 percent [10]. Nobody subtracts the placebo arm. Doing so gives each drug's advantage in percentage points:
- Lamotrigine: +12 on depression, +5 on mania (total 17)
- Lithium: +1 on depression, +14 on mania (total 15)
Near mirror images, with nearly equal totals distributed in opposite directions. The paper states the direction qualitatively — "lamotrigine predominantly effective against depression and lithium predominantly effective against mania" — and never puts a number on either side of it. The two drugs buy about the same total protection and buy it at opposite poles, which is precisely why the choice between them is made on which pole a given patient's illness runs to.
An independent design agrees about lithium. A meta-analysis of five separate randomised trials (770 participants) found lithium's relative risk of relapse was 0.65 for all relapses, 0.62 for manic relapses and 0.72 for depressive relapses — with only the manic interval excluding 1 (0.40–0.95 against 0.49–1.07), the authors calling the depressive protection "smaller and less robust" [11]. Two unrelated study designs, the same asymmetry: lithium protects against the high considerably better than against the low.
Two honest limits. The one-year proportions come from the recently-depressed cohort specifically, so the polarity split may partly reflect index mood; the pooled analysis notes that after adjusting for index mood, "only lithium was superior to placebo for intervention for mania" [9]. And these are stabilised patients who tolerated open-label treatment — 638 of 1,315 — so the trial population is enriched for people who do well on these drugs.
The dose matters, and there is a trial for it. Patients maintained at lithium levels of 0.4–0.6 mmol/L had a relapse risk 2.6 times higher than those at 0.8–1.0 mmol/L, at the cost of more tremor, diarrhoea, urinary frequency, weight gain and metallic taste [12]. Undertreating to reduce side effects has a measurable price.
And stopping is worse than never starting was. Across 14 studies and 257 bipolar I patients, more than 50 percent of new episodes occurred within 10 weeks of stopping an average of 30 months of lithium; the computed time to 50 percent failure of remission was 5.0 months; and the time to 25 percent recurrence of mania was 5.2 times earlier than for depression (2.7 versus 14 months). The authors note the risk of early recurrence "may exceed that predicted by the course of the untreated disorder" [13]. That is the strongest possible argument for not stopping maintenance treatment casually when someone feels well — and, notably, the polarity asymmetry appears here too.
The teaching point. This is an episodic, recurrent illness, so the biggest available gains come from lowering the recurrence rate rather than from treating the current episode faster. Maintenance roughly doubles the time to the next episode; which pole it protects depends on which drug you choose.
Pillar 1: measurement and diagnosis
There is no test, and the diagnosis hinges on a history the patient may not volunteer
Bipolar disorder is diagnosed clinically, from a history of episodes. No blood test, scan or genetic assay makes the diagnosis. And the pivotal element of that history — a past high — is the element least likely to be reported spontaneously, because hypomania is not distressing and is often remembered as a good period.
The consequence is measurable. In a community sample screened with the Mood Disorder Questionnaire, 3.7 percent screened positive for bipolar I or II after adjustment for non-response. Of those, only 19.8 percent said they had ever been given a diagnosis of bipolar disorder; 31.2 percent had been diagnosed with unipolar depression; and 49.0 percent had received neither diagnosis [14]. More people with a positive bipolar screen had been told they had unipolar depression than had been told they had bipolar disorder.
Why this happens is visible in the natural-history data, and it is the most important quantitative fact in this review. Two prospective studies followed patients weekly for over a decade. In bipolar I, patients were symptomatically ill 47.3 percent of weeks across a mean of 12.8 years, of which depressive symptoms occupied 31.9 percent of weeks against 8.9 percent for manic or hypomanic symptoms and 5.9 percent cycling or mixed [15]. In bipolar II, patients were symptomatic 53.9 percent of weeks over a mean 13.4 years, with depressive symptoms at 50.3 percent of weeks against hypomanic at 1.3 percent [16].
The check the arithmetic was never given. These two companion papers, from the same group and a year apart, print those percentages and never form the ratio or compare the two disorders. Doing so:
- Bipolar I: 31.9 / 8.9 = 3.6 weeks depressed for every week high.
- Bipolar II: 50.3 / 1.3 = 38.7 weeks depressed for every week high.
A person with bipolar II spends roughly thirty-nine weeks low for every week high. When they present to a doctor, the overwhelmingly likely state they are in is depression. The misdiagnosis is not carelessness; it is the base rate.
A third, independent dataset says the same. In STEP-BD, 858 patients recovered and were followed 2 years: 48.5 percent had a recurrence, with more than twice as many depressive (34.7 percent) as manic, hypomanic or mixed (13.8 percent) — and the time until 25 percent had a depressive recurrence was 21.4 weeks against 85.0 weeks for a manic one, a four-fold difference [17].
Both illnesses are also mostly subsyndromal. In bipolar I, subsyndromal, minor depressive and hypomanic symptoms combined were nearly three times more frequent than full syndromal episodes (29.9 versus 11.2 percent of weeks), with patients changing symptom status about six times a year and polarity more than three [15]; the bipolar II picture is the same [16]. The textbook image of discrete episodes separated by clean euthymia is not what the weekly data show.
Screening, and what it can and cannot do
The Mood Disorder Questionnaire is the standard brief screen: a score of 7 or more items yielded sensitivity 0.73 and specificity 0.90 against a structured diagnostic interview in 198 mood-clinic outpatients [18]. Six candidate screening instruments have been compared head to head in youth [19]. As with every screen in psychiatry, a positive result is a prompt to take a careful history, not a diagnosis.
The practical instruction that follows from all of the above is simple: in anyone presenting with depression, ask about past highs — periods of reduced need for sleep with increased energy, unusual productivity or spending, racing thoughts, uncharacteristic risk-taking — and ask a family member if you can, because insight into past hypomania is often poor. The case for redefining the boundaries of the spectrum, giving more weight to family history and to antidepressant-induced manic symptoms, has been argued explicitly under the heading of misdiagnosis [20] [21] [22] [23].
What it has to be distinguished from
Unipolar depression is the main one, and the discriminator is the history of highs. Borderline personality disorder produces mood instability, but the shifts are typically hours to days, reactive to interpersonal events, and continuous rather than episodic — against the days-to-months, largely autonomous episodes of bipolar disorder. ADHD shares distractibility, talkativeness and restlessness, but is a stable trait from childhood rather than an episodic departure from a person's baseline. Substance use and medical causes must be excluded; comorbidity between mania and medical and other psychiatric disorders is well described [24].
Comorbidity is the rule, not the exception. Three-quarters of people with bipolar spectrum disorder met criteria for at least one other disorder, anxiety disorders — particularly panic attacks — being the most common [2] [1].
Mood charting — a daily record of mood, sleep and medication — is the practical tool that turns a hard retrospective question into an easier prospective one, and the life-chart method underlies much of the switch literature cited below [25].
Heritability, which is high and often misunderstood
Bipolar disorder is among the most heritable of common psychiatric conditions. A twin study estimated heritability at 85 percent (95% CI 0.73–0.93) narrow and 89 percent broad, with no shared environmental effects detected [26]; a nationwide twin sample found probandwise concordance of 0.43 in monozygotic against 0.06 in dizygotic twins, with a heritability estimate of 0.93 [27]. Family, twin and adoption studies together establish the pattern [28] [29] [30] [31].
But high heritability is not simple inheritance, and the same study makes the point. Fitting a correlated-liability model gave a genetic correlation of 0.65 between mania and depression — and, crucially, about 71 percent of the genetic variance for mania was not shared with depression [26]. Mania has its own genetic architecture. Meanwhile the largest genome-wide association study, 41,917 cases and 371,549 controls, identified 64 loci, enriched in synaptic-signalling and brain-expressed genes and specifically in genes encoding the targets of antipsychotics, calcium-channel blockers, antiepileptics and anaesthetics [32] — with the earlier collaborative analysis having implicated ANK3 and CACNA1C [33]. Sixty-four loci across forty thousand cases is the signature of a polygenic condition, and the genetic overlap with schizophrenia is substantial [34] [35].
Pillar 2: management
Lithium
Lithium is the reference treatment, and the evidence for it is the oldest and in one respect unique.
For maintenance it reduces relapse (RR 0.65 overall, 0.62 manic) [11], with the dose–response and discontinuation findings above [12] [13]. For acute mania it produced at least 50 percent improvement in 49 percent of patients against 25 percent on placebo [36].
The unique claim is about suicide, and it is well replicated. A meta-analysis of 48 randomised trials (6,674 participants) found lithium reduced suicides (odds ratio 0.13, 95% CI 0.03–0.66) and deaths from any cause (0.38, 0.15–0.95) against placebo, without clear benefit for deliberate self-harm [37]. Observational meta-analyses agree on the magnitude: across 31 studies and 85,229 person-years, suicides and attempts were about five times less frequent on lithium (RR 4.91, 3.82–6.31) — roughly an 80 percent reduction — with the attempts-to-suicides ratio rising 2.5-fold, indicating reduced lethality of the acts that did occur [38]; an earlier meta-analysis found suicide 82 percent less frequent during lithium treatment [39] [40] [41]. And in a health-plan cohort, risk of suicide death was 2.7 times higher during divalproex than during lithium treatment after adjustment [42].
The authors of the randomised meta-analysis are careful about mechanism: lithium may act by reducing relapse, but there is evidence it also decreases aggression and possibly impulsivity [37]. This is the one place in psychiatry where a specific drug has a replicated association with reduced mortality, and it is the main reason lithium remains first-line despite being inconvenient.
The inconvenience is real and must be stated. Lithium has a narrow therapeutic index and requires serum-level monitoring, with renal and thyroid function checked periodically; its toxicity profile and side effects have been systematically reviewed [43] [44], and therapeutic drug and cardiovascular monitoring in bipolar disorder is a defined clinical task [45].
Mechanistically it remains incompletely understood, which is unusual for a drug this old. Lithium inhibits glycogen synthase kinase-3 beta, antagonising dopamine-dependent behaviours through an Akt/GSK-3 signalling cascade [46]; it inhibits inositol monophosphatase, depleting free inositol and lowering IP3, which induces autophagy by a route independent of GSK-3β [47]; it has effects on development [48], enhances hippocampal neurogenesis [49], increases the neuroprotective protein Bcl-2 alongside valproate [50], and is associated with increased human brain grey matter [51].
Valproate, carbamazepine and lamotrigine
Valproate is effective in acute mania — 48 percent achieving at least 50 percent improvement against 25 percent on placebo, with efficacy independent of rapid cycling [36] [52] [53] — and has been compared with lithium in a 12-month maintenance trial [54] and tested in combination [55]. It inhibits histone deacetylase 1 directly (IC50 0.4 mM), causing histone hyperacetylation at therapeutic concentrations [56], and also inhibits GSK-3β [57] [58]. Its teratogenicity is the decisive practical constraint: prenatal valproate exposure is associated with autism spectrum disorders and childhood autism [59], and it carries endocrine and metabolic effects in women including polycystic ovaries and hyperandrogenism [60] [61]. It is not a first choice for women who could become pregnant.
Carbamazepine has prophylactic efficacy but in a direct three-arm comparison lithium was more effective for prophylaxis of mania; response rates were 33.3 percent on lithium, 31.4 percent on carbamazepine and 55.2 percent on the combination — a difference that did not reach significance — with the important sub-finding that patients with a history of rapid cycling did poorly on either monotherapy (28.0 and 19.0 percent) and significantly better on the combination (56.3 percent) [62] [63]. The authors' own conclusion is worth keeping: there was "a high incidence of inadequate response to either mood stabilizer or their combination".
Lamotrigine is the depression-side agent. In acute bipolar depression, CGI-improvement response was 51 percent at 200 mg/day and 41 percent at 50 mg/day against 26 percent on placebo [64]. An independent individual-patient meta-analysis of 1,072 participants across five trials found modest overall benefit (RR 1.27 on the HRSD, 1.22 on the MADRS) with a significant interaction by baseline severity — superior to placebo only in those with HRSD above 24 [65] [66]. In maintenance it is the mirror of lithium, as the centerpiece shows, and it has prophylactic effect in rapid cycling [67] [68]. Its constraint is rash, including rare serious rash, which is why titration is slow [69] [70].
Antipsychotics
Second-generation antipsychotics are effective for acute mania — olanzapine produced response in 48.6 percent against 24.2 percent on placebo [71], with aripiprazole [72], risperidone [73] and haloperidol comparisons [74] establishing the class, and combination with lithium or valproate outperforming monotherapy in partial responders [75]. Olanzapine has been compared with divalproex for acute mania and maintenance of remission [76].
Quetiapine is the notable case because it works on the depressive pole, which most antimanic drugs do not. In bipolar I and II depression, response was 58.2 and 57.6 percent at 600 and 300 mg/day against 36.1 percent on placebo, with remission 52.9 against 28.4 percent, and — importantly — treatment-emergent mania was low and no more common than on placebo (3.2 versus 3.9 percent) [77]; a replication found effect sizes of 0.61 and 0.54 with treatment-emergent mania lower than placebo [78]. The olanzapine–fluoxetine combination is also efficacious in bipolar I depression [79] [80].
The cost is metabolic. Antipsychotic-induced weight gain is substantial and differs between agents [81], and the physical-disease risks of antipsychotics, antidepressants and mood stabilisers in this population have been quantified [82]. Comparative efficacy and tolerability across 32 oral antipsychotics is available [83].
Antidepressants: the caution, stated accurately
This is where honesty matters most, because the evidence is genuinely mixed and both the enthusiastic and the alarmed positions overstate it.
The case for caution has three components. First, antidepressant monotherapy raises the risk of mania: in a national registry study, the increased risk of treatment-emergent mania was confined to patients on antidepressant monotherapy (hazard ratio 2.83, 1.12–7.19), while among those also taking a mood stabiliser there was no acute increase (HR 0.79) and a decreased risk at 3–9 months (HR 0.63, 0.42–0.93) [84]. Second, switch rates in naturalistic data are not small: across 228 acute antidepressant trials, threshold switches into hypomania and mania occurred in 11.4 and 7.9 percent of acute trials and 21.8 and 14.9 percent of continuation trials, higher in bipolar I (30.8 percent) than bipolar II (18.6 percent), with venlafaxine showing a much higher switch-to-subthreshold ratio than bupropion or sertraline; only 23.3 percent of patients had a sustained response without a threshold switch [25]. Third, in a clinical-records comparison, manic switching and cycle acceleration occurred only in bipolar depression, and switching was less frequent on mood stabilisers (31.6 versus 84.2 percent) [85] [86] [87] [88].
The case against overstating it is equally real. A systematic review of randomised trials found antidepressants more effective than placebo in short-term bipolar depression, with switching in 3.8 percent on antidepressant against 4.7 percent on placebo — no excess — though tricyclics switched at 10 percent against 3.2 percent for other agents [89]. And STEP-BD, the large effectiveness trial, found that adding a standard antidepressant to a mood stabiliser produced durable recovery in 23.5 percent against 27.3 percent on mood stabiliser plus placebo (P = 0.40) — no benefit — but also no increase in treatment-emergent affective switch [90].
The defensible synthesis, which is what the international consensus reached. The ISBD task force could not endorse broad antidepressant use, noted the gap between wide prescribing and weak evidence, judged SSRIs and bupropion lower-risk for switching than tricyclics and SNRIs, and found mood elevations more frequent and severe in bipolar I than bipolar II [91]. Put plainly: antidepressant monotherapy should be avoided; adding an antidepressant to a mood stabiliser is probably safe and probably does not help much; and the treatments with the best evidence for bipolar depression are quetiapine [77] [78], lamotrigine in more severe depression [65], and the olanzapine–fluoxetine combination [79]. Discontinuing an antidepressant after remission has itself been studied [92], as has adding a second mood stabiliser instead [93].
Psychosocial treatment, which is not optional
Group psychoeducation — teaching patients to recognise their own early warning signs, understand the illness and adhere to treatment — significantly reduced the number of patients relapsing, the number of recurrences per patient and the number and length of hospitalisations, and increased time to depressive, manic, hypomanic and mixed recurrence, in pharmacologically treated bipolar I and II patients [94], with benefit sustained at five years [95] [96]. Teaching patients to identify early symptoms alone has its own randomised trial [97].
Interpersonal and social rhythm therapy targets sleep–wake and daily-routine regularity on the premise that circadian disruption precipitates episodes. Patients assigned to it in the acute phase survived longer without a new episode (P = .01) regardless of maintenance assignment, and had higher social-rhythm regularity — with the ability to increase that regularity associated with reduced recurrence [98] [99]. That is as close as this literature comes to a mechanism you can act on directly.
Family-focused treatment [100] [101] [102] [103], CBT for severe and recurrent bipolar disorder [104] [105], and the STEP-BD comparison in which intensive psychotherapy beat brief collaborative care on total functioning, relationship functioning and life satisfaction over 9 months [106] complete the picture. Cost-effectiveness has been reviewed [107].
ECT
Electroconvulsive therapy remains the most effective acute treatment for severe depression [108] [109] [110] and has a fifty-year record in acute mania [111]. Its use is reserved for severe, refractory, psychotic or life-threatening presentations, and the practical parameters — electrode placement, stimulus intensity and pulse width, which trade efficacy against cognitive effects — are well characterised [112] [113] [114] [115] [116]. Relapse after ECT without continuation treatment is high, which is why continuation pharmacotherapy or continuation ECT is part of the plan rather than an afterthought [117] [118] [119]. Patients' own perspectives on the treatment have been systematically reviewed [120].
Pillar 3: progress
Earlier detection. The misdiagnosis data [14], the screening instruments [18] [19] and the prodromal literature — a 10-year prospective study of prodromal patterns in Amish youth [121], clinical course in children and adolescents with bipolar spectrum disorders [122], and a randomised trial of family-focused therapy in symptomatic youth at risk [123] — together define the current effort: shorten the gap between first depressive presentation and correct diagnosis.
Lithium mechanism and response prediction. GSK-3β [46] [57], inositol monophosphatase and autophagy [47], neurogenesis [49] and Bcl-2 [50] are converging targets, and the therapeutic potential of lithium and valproate beyond bipolar disorder has been reviewed [124]. Predicting who responds to lithium remains unsolved; biological hypotheses and candidate biomarkers have been surveyed without yielding a clinical test [125] [126] [127].
Maintenance evidence continues to be the highest-yield area, and the STEP-BD finding gives it a target: residual mood symptoms at recovery predicted shorter time to recurrence, for both poles independently, suggesting that treating residual symptoms rather than declaring recovery is where relapse reduction is available [17].
Functional recovery lags symptomatic recovery, badly. In the McLean-Harvard first-episode mania study, by 2 years 98 percent achieved syndromal recovery and 72 percent symptomatic recovery — but only 43 percent achieved functional recovery [128]. Symptom scales are not measuring the thing that matters most to patients, and psychosocial disability tracks depressive symptom severity closely [8] [129] [130].
Digital mood monitoring is the natural extension of mood charting, and the substrate supports the rationale rather than the products: the illness changes symptom status about six times a year and polarity more than three [15], most of that movement is subsyndromal, and social-rhythm regularity is both measurable and causally implicated [98]. This substrate contains no trial of a smartphone mood-monitoring application, so this review does not claim one works.
One destigmatising finding worth stating on its own. In long-term follow-up, subsyndromal hypomanic symptoms were not disabling in bipolar II and may even enhance functioning, while depressive symptoms were at least as disabling as manic ones at equivalent severity — and when asymptomatic, patients with bipolar disorder had good psychosocial functioning [8]. The disability in this illness is concentrated in the depression, and the well periods are genuinely well.
Dig deeper in lmmol
Depression is the condition bipolar disorder is mistaken for, and the two reviews are best read together: more people with a positive bipolar screen have been diagnosed with unipolar depression than with bipolar disorder [14], the reason is that bipolar II patients spend roughly thirty-nine weeks depressed for every week high [16], and the treatment consequence is direct — antidepressant monotherapy, appropriate in unipolar depression, nearly triples the risk of treatment-emergent mania here [84]. Anxiety disorders are the commonest comorbidity, present in three-quarters of people with bipolar spectrum disorder, with panic attacks the single most frequent [2] [1]; anxiety also predicts response to acute treatment of bipolar I [131]. Epilepsy shares three drugs with this review — valproate, carbamazepine and lamotrigine all began as anticonvulsants — and shares their liabilities, including valproate's teratogenicity [59] and its endocrine effects [60]. Thyroid disease matters twice over: thyroid dysfunction can mimic mood episodes, and lithium requires thyroid monitoring [44]. Chronic kidney disease is the other monitoring obligation of long-term lithium [43]. Migraine appears at substantially higher rates among people screening positive for bipolar disorder, alongside allergies, asthma and substance misuse [14]. On the molecular side, glycogen synthase kinase-3 beta is inhibited by both lithium and valproate [46] [57], inositol monophosphatase is lithium's other established target [47], histone deacetylase 1 is valproate's [56], and CACNA1C is among the first replicated genome-wide association findings in this illness [33] [32]. The full collection is at health.