出生时辰的「可判别信息量上界」:分辨率从季节调到时辰,信息会不会涨?

Written by

in

Language

Evidence level: B

把出生时间从「季节」细化到「时辰」,分辨率提高约 90 倍;但已测到的时辰级最大效应来自产房排班表,而不是天时——多出来的信息很可能是「谁在值班」的信息。

正文

上一篇《出生时刻的信息量上限》把账算到了季节这一格:出生季节对精神分裂症的 OR ≈ 1.05,折成判别力 AUC ≈ 0.508,折成信息量约 0.0002 比特,与命理判定所需的 10+ 比特隔着 3-4 个数量级。但那篇留了一个洞——命理用的不是季节,是时辰。季节只有 4 格,时辰有 12 格,分辨率高一个数量级。如果信息量确实随分辨率上升,这个洞就可能补上一块。本文直接测量这个上界。

一、先分清「分辨率」能换来什么、不能换来什么。

概率论上,粗化(coarsening)只会丢信息、不会造信息:把 12 个时辰并成 4 个季节,信息量单调不增。所以「时辰 ≥ 季节」是数学上必然的,真正的问题是多多少。参照系可以看基因组:即使动用一个人上百万个位点的多基因评分,对冠心病的判别力 AUC 也只有 0.608、2 型糖尿病 0.607,精神分裂症能到 0.820(PMID 28065900)——用上全部基因组,多数常见病的判别力仍在 0.6 附近。而且 OR 大不等于判别力大:OR = 3 的指标对 AUC 的增量也可能微乎其微(PMID 17309939)。出生时间要谈「信息量跃升」,得先跨过这条基准线。

二、时辰级确实测到了比季节级大一个量级的效应——但方向不对。

加州 1992-1999 年 336 万例出生记录显示:夜间出生(19:00-06:00)的校正后新生儿死亡几率比白天高 12%-16%,这笔超额占了全部新生儿死亡的 9.6%(PMID 16055587)。瑞典 1973-95 年 210 万例自然产中,夜间出生的早新生儿死亡相对风险 RR = 1.31,窒息相关死亡 RR = 1.70,早产儿更明显,且二十年间没有改善(PMID 11739216)。这两个效应量比季节级的 OR 1.05 大了整整一个量级——从信息量角度看,时辰这一格确实装了更多东西。

但机制一查清,方向就变了。2010 年 NICHD 极低出生体重队列(11,137 例)发现出生时机对死亡与神经发育基本无影响,作者直接把结果读成「人力配备是够的」(PMID 20643715)。英格兰 2005-2014 年 605 万例单胎的连锁数据,在校正分娩方式与产程启动方式后,工作时段内外的新生儿死亡率没有差异——所谓的「非工作时段效应」主要由极少数「无产程记录的急诊剖宫产」造成(PMID 37311636)。

结论是硬的:时辰级「信息」的主体不是出生瞬间的物理条件,而是那一小时谁在值班。它可被管理措施消掉——同一国家、同一年代,把人力补上,信号归零。

三、天然对照:非医疗化人群里,时辰由生物学决定,不由排班决定。

斯洛文尼亚 1990-2018 年 55 万例单胎的研究给出了最干净的分解:引产(时机由医务人员班表决定)峰值在上午 10 点;自发宫缩启动的产程峰值在清晨 6-7 点;自发破膜启动的峰值在凌晨 1 点;而且节律分别在 35 周(宫缩型)和 31 周(破膜型)就消失了(PMID 38369585)。作者干脆把引产当作「人力钟」的对照组——这是文献里最明确的一次把两个时钟拆开。

非医疗化人群的数据也一致:马萨诸塞 619,455 例低危自然产只呈现轻微昼夜节律,峰值 11:00-13:00、谷值 23:00-01:00(PMID 10630151);马德里 1887-1892 年 4,599 例产院记录(无人工光照、几乎无产科干预)呈明确清晨峰,且冬至/夏至的峰值位置发生位移,说明光是驱动项(PMID 24789752);阿根廷北部农村 2,421 例的平均方向落在清晨 5:24,作者归因于褪黑素等激素的昼夜节律(PMID 29665070)。

把这几条叠起来看:时辰的真实生物学内容(产程启动的昼夜闸门)是窄的、集中的——大部分自发产程落在午夜到清晨那一段,12 个时辰里先天就有 6-7 个是空档。再加上现代人群里大量时辰是「被安排的」,可用信息被两重压缩。

四、季节这条线上唯一「消不掉」的信号,恰好也不在出生瞬间。

如果出生时间真能留下个体印记,最该留下的地方就是昼夜节律本身。这块数据确实存在:1,584 名大学生中,秋冬出生者晨型更多(PMID 10553957);扩到 3,709 人的余弦拟合显示 MEQ 在 12-1 月出生处最高、6-7 月最低(PMID 12566940);1,591 名成人中,出生时处于 8-10 小时短光周期者晨型更强(PMID 16889896)。

但注意两件事:一是这些是季节/光周期粒度的信号,不是时辰粒度的;二是效应来自发育期的光环境(孕晚期到围产期的光周期),作用窗口在出生之前,不在出生那一秒。也就是说,唯一一条不易被管理措施抹掉、稳定可复制的「出生时间 → 个体差异」通路,恰恰落在季节这条线上,而它的机制是孕期光史,不是出生瞬间的气。

五、这个上界该怎么直接测。

设计上四点:(1)排除排班污染——只在自发产程、且记录了产程启动时刻的子集里分析,或把引产作为显式的「人力钟」协变量;(2)逐粒度对比——把出生时间离散成 4 / 12 / 24 / 144 格,每格算一次互信息(MI,单位比特)与 AUC,看曲线是单调爬升还是早早封顶(互信息估计方法见 PMID 37099536,信息论因果判据见 PMID 18648583);(3)设真实基线——以季节级已测的 OR ≈ 1.05 与光周期 → 时型效应作锚;(4)报告上界而非点估计——用千万级登记数据给出置信上界,直接回答「跃升有没有发生」。

可证伪预测:从 4 格到 12 格,互信息的增量应远小于按格数比例外推的预期——因为非医疗化人群中产程启动高度集中在窄时段,12 格里真正「有人」的格子本来就不足一半。

证据与局限

已验证: 出生时辰(昼夜)与新生儿死亡、窒息相关死亡的关联在大样本中稳定可复制(PMID 16055587、11739216),且机制指向产科人力而非出生瞬间条件(PMID 37311636、20643715);产程启动的昼夜节律与引产的「人力钟」峰分属不同机制(PMID 38369585);非医疗化人群的出生时辰分布集中且受光照驱动(PMID 24789752、29665070、10630151);季节/光周期 → 成人时型关联可复制(PMID 10553957、12566940、16889896)。

部分验证: 「时辰级信息量只比季节级多一点点」是本研究要测的假设,目前只有间接证据(产程启动高度集中、现代时辰被排班内生);互信息逐粒度对比在方法学上可行,但尚未在「出生时间-健康结局」上系统做过;多基因评分 AUC 基准(PMID 28065900)来自不同结局,只能作量级参照。

待验证: 在千万级出生登记 + 结局库(需完整记录出生时刻的国家登记)上直接估计 MI(时辰; 结局) 的置信上界;时辰粒度相对季节粒度的增量是否随年代(医疗化程度上升)单调缩小;跨结局(精神/免疫/代谢)的增量是否一致;非医疗化历史队列(如 19 世纪产院档案)能否提供「无排班污染」的对照上界。

相关阅读

证据等级: B
参考文献: PMID 16055587, PMID 11739216, PMID 37311636, PMID 20643715, PMID 38369585, PMID 10630151, PMID 24789752, PMID 29665070, PMID 10553957, PMID 12566940, PMID 16889896, PMID 28065900, PMID 17309939, PMID 37099536, PMID 18648583


免责说明:本文属传统知识与科学文献的整理、对照与研究设计讨论,仅供理解与参考,不构成医疗建议,不能替代专业诊断与治疗。文中涉及的效应量、频率区间与暴露结论多为研究报告层面的推断,个体情况差异很大;如有身体不适或健康疑虑,请就医并由专业医师判断。

Upper Bound of Discriminable Information in Birth Time: Does Information Increase When Resolving from Season to Hour?

Evidence level: B

Refining birth time from “season” to “hour” increases resolution by approximately 90 times; however, the largest observed hour-level effect comes from hospital scheduling rosters rather than celestial timing—the additional information likely reflects “who is on duty.”

Main Text

The previous article, The Upper Limit of Information at Birth Moment, calculated the account down to the “season” level: birth season shows an odds ratio (OR) ≈ 1.05 for schizophrenia, translating to a discriminative ability AUC ≈ 0.508 and an information content of approximately 0.0002 bits, separated from the ~10+ bits required for metaphysical judgment by 3–4 orders of magnitude. However, that article left an opening—metaphysical practice does not use season; it uses Hour (时辰). With only 4 seasons, there are 12 hours, representing a resolution increase of one order of magnitude. If information content indeed rises with resolution, this gap could be filled. This article directly measures that upper bound.

I. First, distinguish what resolution gains and what it does not gain.

In probability theory, coarsening only loses information; it never creates it: merging 12 hours into 4 seasons results in monotonically non-increasing information. Therefore, “Hour ≥ Season” is mathematically inevitable; the real question is how much more. A reference point can be drawn from genomics: even utilizing polygenic scores across millions of loci in an individual, the discriminative AUC for coronary heart disease is only 0.608, type 2 diabetes 0.607, and schizophrenia reaches 0.820 (PMID 28065900)—even with the entire genome, discriminative ability for most common diseases remains near 0.6. Moreover, a large OR does not equal high discriminative ability: an indicator with OR = 3 may still yield a negligible increment in AUC (PMID 17309939). To discuss an “information leap” for birth time, one must first cross this baseline.

II. Hour-level effects are indeed measured at an order of magnitude larger than season-level—but the direction is wrong.

California birth records from 1992–1999 (3.36 million cases) show that newborns born at night (19:00–06:00) have a corrected mortality risk 12%–16% higher than those born during the day; this excess accounts for 9.6% of all newborn deaths (PMID 16055587). In Sweden, among 2.1 million natural births from 1973–1995, night births showed a relative risk (RR) of 1.31 for early neonatal death, 1.70 for suffocation-related deaths, with more pronounced effects in preterm infants, showing no improvement over two decades (PMID 11739216). These effect sizes are an order of magnitude larger than the season-level OR of 1.05—from an information perspective, the hour level indeed contains more content.

However, once the mechanism is clarified, the direction changes. The NICHD Extremely Low Birthweight cohort (11,137 cases) from 2010 found that birth timing had essentially no effect on mortality or neurodevelopment, with authors interpreting the result as “human staffing is sufficient” (PMID 20643715). Linked data from England covering 6.05 million singleton births from 2005–2014 showed no difference in neonatal mortality between working and non-working hours after adjusting for mode of delivery and induction timing—the so-called “non-working hour effect” was primarily caused by a small number of “emergency cesarean sections without labor records” (PMID 37311636).

The conclusion is firm: the subject of hour-level “information” is not the physical conditions at birth, but who is on duty during that hour. It can be eliminated by management measures—within the same country and decade, supplementing staffing causes the signal to vanish.

III. Natural Control: In non-medicalized populations, hour is determined biologically, not by scheduling.

A Slovenian study of 550,000 singleton births from 1990–2018 provided the cleanest decomposition: induced labor (timing determined by medical staff rosters) peaks at 10:00 AM; labor onset from spontaneous contractions peaks between 6:00–7:00 AM; and rupture of membranes initiated spontaneously peaks at 1:00 AM, with rhythms disappearing by 35 weeks (contraction type) and 31 weeks (membrane type), respectively (PMID 38369585). The authors explicitly treat induced labor as a control group for the “human clock”—the clearest instance in literature of separating two clocks.

Non-medicalized population data is consistent: Massachusetts 619,455 low-risk natural births show only a mild circadian rhythm, peaking 11:00–13:00 and troughing 23:00–01:00 (PMID 10630151); Madrid records from 1887–1892 (4,599 cases, no artificial light, minimal obstetric intervention) show a clear morning peak with peak position shifting between solstices, indicating light as the driver (PMID 24789752); rural northern Argentina with 2,421 cases shows an average direction at 5:24 AM, attributed to circadian rhythms of hormones like melatonin (PMID 29665070).

Viewing these together: the true biological content of the hour (the circadian gate for labor onset) is narrow and concentrated—most spontaneous labor falls within the midnight-to-morning window, leaving 6–7 of the 12 hours naturally empty. Plus, in modern populations, a large portion of hours are “scheduled,” resulting in double compression of available information.

IV. The only signal that cannot be erased on the season line does not occur at birth itself.

If birth time truly leaves an individual imprint, the most logical place would be the circadian rhythm itself. Such data exists: among 1,584 college students, those born in autumn/winter are more morning-type (PMID 10553957); expanding to 3,709 individuals via cosine fitting shows MEQ peaks in December–January and troughs in June–July (PMID 12566940); among 1,591 adults, those born during short light cycles (8–10 hours) show stronger morning types (PMID 16889896).

But note two things: First, these are signals at the season/light cycle granularity, not hour granularity; second, the effect stems from the light environment during development (late pregnancy to perinatal light cycles), with the action window occurring before birth, not at the second of birth. Thus, the only pathway from “birth time → individual difference” that is resistant to management measures, stable, and replicable falls on the season line, with its mechanism being pregnancy light history, not the air at birth.

V. How to directly measure this upper bound.

Design involves four points: (1) Exclude scheduling contamination—analyze only subsets of spontaneous labor with recorded onset times, or treat induced labor as an explicit “human clock” covariate; (2) Compare across granularities—discretize birth time into 4 / 12 / 24 / 144 bins, calculating mutual information (MI, in bits) and AUC for each to observe if the curve climbs monotonically or caps early (MI estimation methods see PMID 37099536, causal information theory criteria see PMID 18648583); (3) Establish a true baseline—anchor using the season-level OR ≈ 1.05 and light cycle → chronotype effects; (4) Report upper bounds rather than point estimates—use registry data at the million-plus level to provide confidence upper bounds, directly answering “did a leap occur.”

Falsifiable prediction: The incremental mutual information from 4 bins to 12 bins should be far smaller than extrapolated by bin proportion—because in non-medicalized populations, labor onset is highly concentrated in narrow time windows, and fewer than half of the 12 hours are naturally “occupied.”

Evidence and Limitations

Verified: The association between birth hour (circadian) and neonatal death, suffocation-related death is stable and replicable in large samples (PMID 16055587, 11739216), with mechanisms pointing to obstetric staffing rather than birth conditions (PMID 37311636, 20643715); circadian rhythms of labor onset and induced labor’s “human clock” peak belong to different mechanisms (PMID 38369585); birth hour distribution in non-medicalized populations is concentrated and driven by light (PMID 24789752, 29665070, 10630151); season/light cycle → adult chronotype associations are replicable (PMID 10553957, 12566940, 16889896).

Partially Verified: “Hour-level information content is only slightly higher than season-level” is the hypothesis this study tests; currently only indirect evidence exists (high concentration of labor onset, modern hours endogenously scheduled by rosters); cross-granularity MI comparison is methodologically feasible but not yet systematically applied to “birth time–health outcome”; polygenic score AUC benchmarks (PMID 28065900) come from different outcomes and serve only as order-of-magnitude references.

To Be Verified: Directly estimate the confidence upper bound of MI(Hour; Outcome) using million-plus birth registries with outcome databases (requiring complete recording of birth moments in national registries); whether the incremental increase relative to season granularity monotonically shrinks with era (increasing medicalization); whether increments are consistent across outcomes (psychiatric/immunometabolic); whether non-medicalized historical cohorts (e.g., 19th-century hospital archives) can provide a “no-scheduling-contamination” upper bound control.

Related Reading

Evidence Level: B
References: PMID 16055587, PMID 11739216, PMID 37311636, PMID 20643715, PMID 38369585, PMID 10630151, PMID 24789752, PMID 29665070, PMID 10553957, PMID 12566940, PMID 16889896, PMID 28065900, PMID 17309939, PMID 37099536, PMID 18648583


Disclaimer: This article compiles traditional knowledge and scientific literature for comparison and research-design discussion only. It is not medical advice and cannot replace professional diagnosis or treatment. Effect sizes, frequency ranges and exposure conclusions cited here are largely inferences at the research-report level, and individual circumstances vary widely. If you have any discomfort or health concerns, please consult a physician.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *