One major problem of preventive psychiatry is the limited predictive strength of all known risk factors for mental disorders, meaning that most of the individuals who are judged to be at risk have only a small chance of developing a mental disorder within the next period of their lives. Fusar‐Poli et al¹ have produced an excellent overview of the current state of preventive psychiatry, and they refer to this problem several times. However, we think this is a key issue that deserves more exploration, because it can also give directions for how the prevention field can move forward.

The problem of the low predictive strength of risk factors is partly related to the different priorities of epidemiological research and prevention science. In epidemiological research, the relative risk (RR) or the odds ratio (OR) is often the main indicator describing the strength of the association between a risk factor and a health outcome. However, these indicators only have limited value for prevention science.

For example, if the incidence of a mental disorder in the next year is 1% of the population, and the RR of a group at risk is 4, that means that 4% of this high‐risk group will develop the disorder instead of only 1% in the general population. Epidemiological researchers usually stop when they find a (significant) RR of 4, because this indicates a clear high‐risk group. However, this is not enough for prevention science. A preventive intervention for a group with 4% risk (instead of 1%) still means that almost all people with this risk factor (96%) will not develop the disorder. Suppose that a preventive intervention can reduce this risk from 4% to 2%. That means that, of the 100 high‐risk participants in the intervention, 96 would not develop the disorder anyway and, of the 4 who would, only two will benefit from the preventive effect. This is neither cost‐effective nor ethical.

Unfortunately, even though high RRs and ORs are often found in epidemiological research, almost all risk factors in mental health suffer from a low predictive strength. Having a parent with a depressive disorder is often given as an example of a group with an exceptionally high risk. One study even indicates that 50% of these children will develop a depression by the age of 20², which is much larger than any other risk factor for mental disorders. But, from the perspective of preventive interventions, even such an elevated incidence rate is still problematic. Suppose that the development of depression starts at the age of 12 and is evenly divided over the subsequent 8 years. This means that every year still only about 6% of these children will develop depression. Offering a preventive intervention to a group in which 94% will not develop the disorder in the following year is still problematic.

Screening for high‐risk groups has com­parable problems. For example, testing positive for high risk for psychosis has been found to be associated with a 6% lifetime risk of actually developing psychosis³. This means that 94% of those who score positive will not develop psychosis in their lifetime, and it can be disputed whether preventive interventions should be considered in these cases⁴.

So, from the perspective of preventive interventions, RRs and ORs are clearly not sufficient as indicators of risk. An absolute risk of developing a disorder within a reasonable time frame would be a better indicator. In addition, we need to take the prevalence of the risk factor in the population into account (exposure prevalence), because that indicates the size of the population that will have to be given the intervention.

For example, it is known that women have a higher chance of developing a depressive disorder, but intervening in half of the population is simply not feasible nor cost‐efficient (apart from all ethical issues). On the other hand, an intervention in a small group (i.e., with a small exposure prevalence) and a high risk may be useful for the individual participants, but it will not have a large impact on the incidence of a disorder in the general population. This implies that, from the perspective of preventive interventions, we need to identify a population with a modest prevalence (because otherwise the cost of intervening is too high), but this population should be responsible for as many new cases as possible, meaning that the absolute risk is as high as possible in this group.

Finally, preventive interventions should reduce the incidence of the disorder in the population as much as possible. From this perspective, the weak predictive power of most risk indicators is also problematic, because the lower the incidence rate in the population, the larger randomized trials need to be, in order to have sufficient statistical power to be able to show a significant reduction of the incidence⁵. For example, if we were able to identify a high‐risk group with 25% incidence in the next year and we had an intervention that is capable to reduce the incidence to 17%, we would need a trial of about 1,000 participants (assuming an alpha of 0.05, 80% power and 20% attrition)⁵.

How can this problem of the low predictive power of most risk factors be solved? One possible solution is to focus on combinations of risk factors, that identify groups that are as small as possible but are at the same time responsible for as many incident cases as possible. For example, in one study among older adults, we found that those with sub‐threshold depression, functional limitations, a small social network and female gender were 8% of the population, but they explained 24% of the new incident cases of depression⁶.

A related solution is to develop prediction tools to identify individuals with a much increased risk for developing mental disorders. The PredictD method has been studied in several large European epidemiological studies⁷. A comparable method has been developed in the US⁸. Based on well‐established predictors for the development of depression, these methods calculate the exact personal risk to develop a depressive disorder in the coming year. Unfortunately, these methods do not solve the problem of the low specificity of known risk factors¹. However, the digitalization of our societies and the progress in epidemiology has resulted in large datasets which may improve such approaches with machine learning techniques.

In addition to the identification of high‐risk groups with greater certainty, we also need better interventions. The impact of preventive interventions not only depends on the absolute risk in the target group, but also on their ability to reduce that risk. Some strategies may strengthen the effects of interventions. For example, by focusing on multiple disorders instead of only one, the absolute risk in the target group may be higher and the effects could be demonstrated easier in prevention trials⁹. Stepped care approaches, in which at‐risk people are followed over time, may also improve outcome, although that has not been confirmed in all studies.

We conclude that the predictive strength of most risk factors for the development of mental disorders is low and the identification of populations at ultra‐high risk is key to the further development of preventive psy­chiatry.

预防精神病学的一个主要问题是，所有已知的精神疾病风险因素的预测强度有限，这意味着大多数被判断为处于危险之中的个体在其下一个生命周期内发展为精神疾病的机会很小。 。Fusar-Poli等人¹对当前的预防精神病学状态进行了很好的概述，他们多次提到了这个问题。但是，我们认为这是一个值得进一步探讨的关键问题，因为它也可以为预防领域的前进方向提供指导。

危险因素的预测强度低的问题部分与流行病学研究和预防科学的不同优先级有关。在流行病学研究中，相对风险（RR）或优势比（OR）通常是描述风险因素与健康结果之间关联强度的主要指标。但是，这些指标仅对预防科学具有有限的价值。

例如，如果第二年精神障碍的发生率占总人口的1％，处于危险中的人群的RR为4，则意味着该高危人群中有4％会发展为精神障碍，而不仅仅是在总人口中占1％。流行病学研究人员通常在RR（显着）为4时停止，因为这表明存在明显的高危人群。但是，这还不足以预防科学。对风险为4％（而不是1％）的人群进行预防性干预仍然意味着，几乎所有具有该风险因素的人（96％）都不会发展为该疾病。假设预防性干预可以将这种风险从4％降低到2％。这意味着，在干预的100名高风险参与者中，有96名无论如何都不会发展为该疾病，而在4名参与者中，只有2名受益于这种预防作用。

不幸的是，即使在流行病学研究中经常发现较高的RR和OR，但几乎所有心理健康风险因素的预测强度都较低。父母通常会患有抑郁症，这是高风险人群的一个例子。一项研究甚至表明，这些儿童中有50％会在20岁之前发展为抑郁症²，这比任何其他精神障碍风险因素大得多。但是，从预防干预的角度来看，即使如此高的发病率仍然是有问题的。假设抑郁症的发生在12岁时开始，并在随后的8年中平均分配。这意味着每年仍然只有约6％的儿童会患上抑郁症。向94％不会在第二年发展为该疾病的人群提供预防性干预仍然存在问题。

对高危人群进行筛查存在类似的问题。例如，已发现对精神病的高风险进行阳性检测与实际患上精神病的6％终生风险相关[ ^3]。这意味着94％的得分为阳性的人一生中不会发展为精神病，因此在这种情况下是否应考虑采取预防性干预措施存在争议⁴。

因此，从预防干预的角度来看，RR和OR显然不足以作为风险指标。在合理的时间内出现疾病的绝对风险将是一个更好的指标。此外，我们需要考虑人群中危险因素的患病率（暴露患病率），因为这表明必须采取干预措施的人群规模。

例如，众所周知，女性患抑郁症的机会更高，但是干预一半的人口根本不可行，也不具成本效益（除了所有道德问题）。另一方面，对一小部分人群（即接触率低）和高风险的干预措施可能对个体参与者有用，但不会对一般人群的疾病发生率产生重大影响。这意味着，从预防干预的角度来看，我们需要确定一个患病率较低的人群（因为否则干预的成本太高），但是该人群应负责尽可能多的新病例，这意味着该组中的绝对风险尽可能高。

最后，预防性干预措施应尽可能减少人群中疾病的发生。从这个角度来看，大多数风险指标的预测能力较弱也是一个问题，因为人群的发生率越低，就需要进行更大的随机试验，以便具有足够的统计能力以显示出显着降低发生率⁵。例如，如果我们能够在明年确定高危人群的发病率为25％，并且我们采取的干预措施能够将发病率降低到17％，那么我们将需要约1,000名参与者的试验（假设α为0.05，功率为80％，损耗为20％）⁵。

如何解决大多数风险因素的低预测能力问题？一种可能的解决方案是集中于风险因素的组合，以识别尽可能小的小组，但同时负责尽可能多的事件案例。例如，在一项针对老年人的研究中，我们发现患有亚阈以下抑郁症，功能受限，社交网络较小和女性的人占总人口的8％，但他们解释了24％的新发生的抑郁症病例⁶。

一个相关的解决方案是开发预测工具，以识别罹患精神障碍的风险大大增加的个体。欧洲几项大型流行病学研究都对PredictD方法进行了研究⁷。美国已开发出一种类似的方法⁸。这些方法基于成熟的抑郁症预测指标，可以计算出来年患抑郁症的确切个人风险。不幸的是，这些方法不能解决已知风险因素^1的特异性低的问题。但是，我们社会的数字化和流行病学的发展导致了庞大的数据集，这些数据集可能会通过机器学习技术来改善这种方法。

除了更确定地确定高风险人群外，我们还需要更好的干预措施。预防性干预措施的影响不仅取决于目标人群的绝对风险，还取决于其降低风险的能力。一些策略可能会增强干预措施的效果。例如，通过关注多种疾病而不是仅关注一种疾病，目标人群的绝对风险可能更高，并且在预防试验中可以更容易地证明其效果⁹。尽管并未在所有研究中均得到证实，但逐步护理的方法可能会改善预后，在这种方法中，高危人群会随着时间的流逝而受到关注。

我们得出的结论是，大多数风险因素对精神疾病发展的预测强度很低，识别超高风险人群是预防精神病学进一步发展的关键。