<p>A major challenge in multimorbid aging is understanding how diseases co-occur and identifying high-risk groups for accelerated disease development, but to date associations in the relative onset acceleration of disease diagnoses have not been used to characterize disease patterns. This study presents the development and evaluation of a neural network Cox model for predicting onset acceleration risk for age-associated conditions, using demographic, anthropomorphic, imaging, and blood biomarker traits from 60,396 individuals and 218,530 outcome events from the UK Biobank. Risk prediction was evaluated with Harrell’s concordance index (C-index). The model performed well on internal (C-index <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$0.6830 \pm 0.0902$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>0.6830</mn> <mo>±</mo> <mn>0.0902</mn> </mrow> </math></EquationSource> </InlineEquation>, <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$n=8,931$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>n</mi> <mo>=</mo> <mn>8</mn> <mo>,</mo> <mn>931</mn> </mrow> </math></EquationSource> </InlineEquation>) and external (C-index <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$0.6461 \pm 0.1264$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>0.6461</mn> <mo>±</mo> <mn>0.1264</mn> </mrow> </math></EquationSource> </InlineEquation>, <InlineEquation ID="IEq4"> <EquationSource Format="TEX">$n=855$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>n</mi> <mo>=</mo> <mn>855</mn> </mrow> </math></EquationSource> </InlineEquation>) test sets, attaining C-index <InlineEquation ID="IEq5"> <EquationSource Format="TEX">$\ge 0.6$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mo>≥</mo> <mn>0.6</mn> </mrow> </math></EquationSource> </InlineEquation> on 38 out of 47 (<InlineEquation ID="IEq6"> <EquationSource Format="TEX">$80.9\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>80.9</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>) conditions. Inclusion of body composition and blood biomarker input traits was independently important for predictive performance. Kaplan-Meier curves for predicted risk quartiles (log-rank <InlineEquation ID="IEq7"> <EquationSource Format="TEX">$p \le 1.16E-16$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>p</mi> <mo>≤</mo> <mn>1.16</mn> <mi>E</mi> <mo>-</mo> <mn>16</mn> </mrow> </math></EquationSource> </InlineEquation>) indicated robust stratification of individuals into high and low risk groups. Analysis of risk quartiles revealed cardiometabolic, vascular-neuropsychiatric, and digestive-neuropsychiatric disease clusters with strong statistically significant inter-correlated onset acceleration (<InlineEquation ID="IEq8"> <EquationSource Format="TEX">$r \ge 0.6$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>r</mi> <mo>≥</mo> <mn>0.6</mn> </mrow> </math></EquationSource> </InlineEquation>, <InlineEquation ID="IEq9"> <EquationSource Format="TEX">$p \le 3.46E-5$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>p</mi> <mo>≤</mo> <mn>3.46</mn> <mi>E</mi> <mo>-</mo> <mn>5</mn> </mrow> </math></EquationSource> </InlineEquation>), while 13 and 19 conditions were strongly associated with onset acceleration of all-cause mortality and all-cause morbidity, respectively. In prognostic survival analysis, the proportional hazards assumption was met (Schoenfeld residual <InlineEquation ID="IEq10"> <EquationSource Format="TEX">$p > 0.05$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>p</mi> <mo>></mo> <mn>0.05</mn> </mrow> </math></EquationSource> </InlineEquation>) in 435 out of 435 or 100% (1238 out of 1334 or 92.8%) of cases across outcomes, <InlineEquation ID="IEq11"> <EquationSource Format="TEX">$aHR= 6.11 \pm 9.00$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>a</mi> <mi>H</mi> <mi>R</mi> <mo>=</mo> <mn>6.11</mn> <mo>±</mo> <mn>9.00</mn> </mrow> </math></EquationSource> </InlineEquation> (<InlineEquation ID="IEq12"> <EquationSource Format="TEX">$aHR = 3.67 \pm 5.78$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>a</mi> <mi>H</mi> <mi>R</mi> <mo>=</mo> <mn>3.67</mn> <mo>±</mo> <mn>5.78</mn> </mrow> </math></EquationSource> </InlineEquation>) with (without) Bonferroni correction. The neural architecture of OnsetNet was interpreted with saliency analysis, and several significant body composition and blood biomarkers were identified. The results demonstrate that neural network survival models are able to estimate prognostically informative onset acceleration risk, which could be used to improve understanding of synchronicity in the onset of age-associated diseases and reprioritize patients based on disease-specific risk.</p>

Deep learning predicts onset acceleration of 38 age-associated diseases from blood and body composition biomarkers in the UK Biobank

期刊论文
出版模式：

开放获取
发表日期：

2025年6月27日

查看全文

Mica Xu Ji,
Marjola Thanaj,
Léna Nehale-Ezzine,
Brandon Whitcher,
E. Louise Thomas,
Jimmy D. Bell

摘要

A major challenge in multimorbid aging is understanding how diseases co-occur and identifying high-risk groups for accelerated disease development, but to date associations in the relative onset acceleration of disease diagnoses have not been used to characterize disease patterns. This study presents the development and evaluation of a neural network Cox model for predicting onset acceleration risk for age-associated conditions, using demographic, anthropomorphic, imaging, and blood biomarker traits from 60,396 individuals and 218,530 outcome events from the UK Biobank. Risk prediction was evaluated with Harrell’s concordance index (C-index). The model performed well on internal (C-index $0.6830 \pm 0.0902$ $0.6830 \pm 0.0902$ , $n=8,931$ $n = 8, 931$ ) and external (C-index $0.6461 \pm 0.1264$ $0.6461 \pm 0.1264$ , $n=855$ $n = 855$ ) test sets, attaining C-index $\ge 0.6$ $\geq 0.6$ on 38 out of 47 ( $80.9\%$ $80.9 %$ ) conditions. Inclusion of body composition and blood biomarker input traits was independently important for predictive performance. Kaplan-Meier curves for predicted risk quartiles (log-rank $p \le 1.16E-16$ $p \leq 1.16 E - 16$ ) indicated robust stratification of individuals into high and low risk groups. Analysis of risk quartiles revealed cardiometabolic, vascular-neuropsychiatric, and digestive-neuropsychiatric disease clusters with strong statistically significant inter-correlated onset acceleration ( $r \ge 0.6$ $r \geq 0.6$ , $p \le 3.46E-5$ $p \leq 3.46 E - 5$ ), while 13 and 19 conditions were strongly associated with onset acceleration of all-cause mortality and all-cause morbidity, respectively. In prognostic survival analysis, the proportional hazards assumption was met (Schoenfeld residual $p > 0.05$ $p > 0.05$ ) in 435 out of 435 or 100% (1238 out of 1334 or 92.8%) of cases across outcomes, $aHR= 6.11 \pm 9.00$ $a H R = 6.11 \pm 9.00$ ( $aHR = 3.67 \pm 5.78$ $a H R = 3.67 \pm 5.78$ ) with (without) Bonferroni correction. The neural architecture of OnsetNet was interpreted with saliency analysis, and several significant body composition and blood biomarkers were identified. The results demonstrate that neural network survival models are able to estimate prognostically informative onset acceleration risk, which could be used to improve understanding of synchronicity in the onset of age-associated diseases and reprioritize patients based on disease-specific risk.