Deep survival modelling to predict future cognitive impairment in unimpaired adults

Abstract

Background:

Predicting Alzheimer’s disease (AD)-related cognitive impairment (CI) among cognitively normal (CN) adults enables meaningful disease modification through early intervention and enrichment of clinical trials.

Methods:

A deep survival model is trained to predict CI conversion risk in 1,415 CN adults from the National Alzheimer’s Coordinating Center. Converters’ (N=212) and non-converters’ (N=1,203) baseline clinical measures and magnetic resonance images are used to estimate their conversion probability up to 22 years after baseline observation.

Results:

After 20-fold cross-validation, the model predicts conversion probability with a c-index of 0.88, and classification accuracy of 75% and AUC ROC of 0.89, outperforming previous machine learning models.

Conclusions:

This is one of few studies on the important challenge of predicting future CI among unimpaired subjects. Deep survival modelling can improve the identification of preclinical AD and suggests that uncertainty in AD risk estimation is due to potentially modifiable lifestyle factors.

1. Background

Identifying patients at risk of cognitive impairment (CI) is essential for reducing the burden of symptomatic Alzheimer’s disease (AD)¹, which is characterized by advanced, irreversible neurodegeneration². To enable meaningful AD modification, therapeutic intervention must be initiated during the preclinical/prodromal stages of the disease³. Early identification of future CI converters among cognitively normal (CN) individuals is also crucial for enriching clinical trial samples with asymptomatic study participants – i.e., those with the underlying disease but no observable symptoms⁴. Currently, there are no routine evaluations that can reliably identify individuals at risk of CI.

Machine learning (ML) can enhance prognostication of preclinical AD⁵ by identifying 68%−71% of future CI converters using socio-demographics, cognitive test scores, and medical histories^6,7. In clinical settings, survival analysis has a considerable advantage over classifiers because they enable subject-specific risk estimation. Classifiers only provide binary outcome predictions; on the other hand, survival models can additionally enable conversion outcome prediction across a clinician-defined time window following observation, while also estimating risk over time at individual level⁸. ML classifiers require the exclusion of censored data (i.e., information from subjects who were lost to follow-up before conversion was detected). Thus, they typically predict conversion only within a relatively short (2- to 3-year) window after observation⁸. In contrast, survival analysis can estimate CI conversion probability over longer time horizons by accommodating censored individuals⁹.

Deep survival models (DSMs) can predict time to CI conversion using complex, nonlinear transformations of clinical and neuroimaging measures, offering alternatives to traditional survival analysis^10–12. For example, Mirabnahrazam et al.¹³ and Nakagawa et al.⁸ utilize DSMs to predict future AD diagnosis in a combined pool of CN and mild CI (MCI) participants, achieving concordance indexes (c-indexes) of 0.83 and 0.84, respectively. The c-index evaluates a model’s ability to correctly order individuals based on their predicted conversion risk scores, quantifying the discrimination between those who experience conversion sooner versus later. However, prognosis within training samples that include both MCI patients and CN subjects is problematic, as persons with MCI are already at higher risk of AD and death¹⁴.

Very few studies investigate CI conversion within baseline samples containing only CN subjects. For example, Li et al.¹⁵ constructed a survival model with a c-index of 0.66 to predict CN-to-MCI conversion using longitudinal clinical and MRI data. Khajehpiri et al.¹⁶ trained a boosted Cox proportional hazards model and predicted CN-to-MCI conversion with a c-index of 0.73. These authors, however, omitted the inclusion of magnetic resonance image (MRI) features beyond hippocampal atrophy as potential CI predictors.

Structural brain changes occur before CI diagnosis^1,17,18, but brain volume (BV) alone does not provide a sensitive or specific measure of proximity to AD. By contrast, our research has found that brain age (BA), the biological (neuroanatomic) age of the brain estimated from T₁-weighted MRIs, correlates with time to CI conversion¹⁹. BA, typically estimated using deep neural networks, conveys the difference between a person’s chronological age (CA) and her/his neuroanatomic age. BA saliency is an explainable artificial intelligence measure of how important each brain location is to the deep neural network when estimating BA. Thus, saliency indexes the relative utility of neuroanatomic features when estimating BA highlighting key structures involved in neuroanatomic aging²⁰.

In this study, we train a DSM¹¹ on clinical and neuroimaging measures to predict CI conversion among CN adults imaged by the National Alzheimer’s Coordinating Center (NACC). To relate neuroanatomical profiles to conversion prediction probabilities, we include both regional volumes as well as BA measures in the DSM. The DSM’s ability to predict survival times and distinguish between converters and non-converters is assessed and benchmarked against seven ML classifiers trained on the same data. We identify variables most predictive of CI conversion and compare conversion risk across clinical subgroups based on socio-demographics, medical histories, and clinical and cognitive factors. This work outperforms previous efforts to identify subjects at CI risk before symptom onset.

2. Methods

2.1. Participants

Data were obtained for 4194 adults from NACC (Supplementary Methods 1.1). Procedures were performed following ethical standards of the 1964 Helsinki Declaration and its later amendments, the US Code of Federal Regulations (45 CFR 46), and with approval from the Institutional Review Board of the University of Southern California. Written informed consent was obtained from all participants (or guardians of participants).

Participants were included if they had at least one T₁-weighted MRI and neuropsychological visit (i.e., baseline) and at least one further neuropsychological visit (i.e., follow-up; N = 1491; 67.54% female; CA range: 21 to 100 years (y); CA mean ± SD = 69.06 ± 10.67 y). CI conversion was defined as a diagnosis of either MCI or AD within the study timeframe based on consensus criteria from the National Institute of Neurological and Communicative Disorders and Stroke, the Alzheimer’s Disease and Related Disorders Association, and the NIA & Alzheimer’s Association (NIA-AA). We excluded subjects with Parkinson’s disease, Lewy body dementia, vascular dementia, multi-infarct dementia, Huntington’s disease, normal pressure hydrocephalus, brain tumor, progressive supranuclear palsy, seizure disorder, subdural hematoma, multiple sclerosis, stroke, other neurological abnormality, or with history of significant head trauma followed by persistent neurologic defaults or structural brain abnormalities.

At baseline, all converters and non-converters were CN. A subject was considered a converter if her/his cognitive status changed to ‘CI’ at follow-up. Converters who ‘reverted’ (i.e., were designated as CN at a visit after their MCI or AD diagnosis) were excluded (N = 76, 26% of all converters). Non-converters remained CN at a follow-up. Non-converters with less than 2.5 years of data available (i.e., follow-up occurred less than 2.5 years after baseline) were excluded.

2.1.1. Censoring

Let t be a survival model’s time-to-event, i.e., the number of days between baseline (initial MRI) and follow-up. This is the variable whose value is predicted by the DSM for participants with uncensored data. Converters’ follow-up visit was defined as the visit where conversion from CN to MCI or AD was first recorded; thus, t is the number of days from initial MRI to CI diagnosis (time to conversion). Subjects who did not convert during their time in the study (i.e., non-converters) were defined as being right-censored, whereby conversion may happen after a certain time, but the exact time is unknown. Their conversion date is thought to occur after the conclusion of the study, if at all. In their case, t is the number of days from initial MRI to the last visit on record.

2.2. Clinical variables

The NACC database contains over 1195 data elements, including clinical measures relating to subjects’ demographics, family history, health history, physical health, and neurocognitive status (interviews, questionnaires, cognitive batteries, and neurological symptoms). We selected 61 variables based on findings from previous studies that identified NACC variables predictive of future CI^6,19,21. Supplementary Table 1 lists each variable, how it was measured, its data type, and measurement scale. Variables were categorized as socio-demographic (e.g.,CA at baseline, sex assigned at birth, years of education), medical (e.g., body mass index, smoking status, history of heart attack or stroke), mental health and daily life functioning (e.g., functional activities questionnaire, geriatric depression scale), clinical (e.g., clinical dementia rating, decline in cognition, judgement, and/or visuospatial function), biomarker (e.g., APOE genotype, amyloid PET and/or CSF status), or cognitive (e.g., verbal fluency, trails A & B). Cognitive tests quantifying the same cognitive function (i.e., the Boston naming test and the multilingual naming test, and Craft story 21 and the logical memory test) were combined and rescored as percentages of the total number of possible correct responses (see¹⁹ for details).

2.3. Brain volumes and BA saliencies

2.3.1. T₁-weighted MRI acquisition and processing

MRI acquisition protocols vary across the research centers included in this study. De-identified T₁-weighted MRIs were submitted voluntarily between 2005 and March 2023. Quality control by visual assessment was performed before and after Freesurfer (https://surfer.nmr.mgh.harvard.edu/, version 6.0.0) processing, which included removal of non-brain tissues, transformation into Talairach space, intensity normalization, segmentation into cortical/subcortical structures, surface processing, and topology correction. Structure-level volumetrics were computed for 185 head compartments according to the Destrieux brain atlas²² (148 cortical, 19 subcortical, 9 white matter, and 9 cerebrospinal fluid (CSF) compartments; see Supplementary Table 2). After allometric scaling (see Supplementary Methods 1.2), the total volumes of cortical grey matter, subcortical grey matter, white matter, and CSF were computed using the respective sums of relevant sub-volumes. For example, total cortical grey matter volume was computed as the sum of 148 regional cortical volumes.

2.3.3. BA and BA saliency

Global BA was estimated using a 3-dimensional convolutional neural network (3D-CNN) which produces accurate and generalizable estimates in adults^20,23. Details on model architecture, training procedure, and performance evaluation are described elsewhere²⁰ and in Supplementary Methods 1.3. Outputs of the model are BA, age gap (AG, where ${AG}_i={BA}_i-{CA}_i$ for each subject i), and BA saliency maps. Saliency maps are brain maps organized topographically according to how salient (i.e., important) anatomic locations are to the 3D-CNN when estimating BA (see Supplementary Methods 1.3). Such maps reveal neuroanatomic patterns of aging. Brain structures more useful during BA estimation have higher saliency, and those less important have lower saliency. One can consider brain structures with higher saliency to be more entrained to the anatomic aging process, in the sense that such structures highlight key features of anatomic brain aging²⁰. We computed average BA saliency probability (hereinafter referred to simply as saliency) for each FreeSurfer atlas brain structure.

2.4. DSM

DSMs distinguish between censored and non-censored data through a loss function that integrates both observed event times and censored data to estimate survival probabilities at various time points. This DSM implements a fully parametric model for survival analysis designed to handle time-to-event data with time-varying covariates, as described elsewhere¹¹ (see Supplementary Methods 1.4). We used the DSM to estimate survival probability (i.e., the probability that an individual will survive beyond a specific time point t) in our sample. Hereinafter, ‘survival probability’ refers to the probability that a subject or group of subjects does not convert within the time frame of the study.

We applied 20-fold cross-validation, with each fold split into 95% training and 5% test data. Survival probabilities and hazard rates were predicted for each subject across t = 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, and 22 years (spanning all subjects’ minimum to maximum observation window, i.e., time in the study). Hazard rates represent the instantaneous rate of conversion and were calculated using the formula $h\left(t\right)=-S\prime \left(t\right)/S\left(t\right)$, where $S\left(t\right)$ is the survival probability at time t and $S\prime \left(t\right)$ is its derivative (i.e., the rate of change from baseline until t). Survival performance was benchmarked on test sets using cumulative dynamic area under the receiver operating curves (AUCROCs), integrated Brier scores, and c-indexes both with and without inverse probability of censoring weighting (IPCW) adjustment (Supplementary Methods 1.5)²⁴. The c-index inherently depends on t and is not directly proportional to the number of subjects whose risk has been accurately predicted²⁵. For this reason, we also examined sensitivity and specificity, by identifying as converters subjects whose survival probability dropped below 50%. At each timepoint t, subjects were classified as converters if their likelihood of survival beyond timepoint t was less than 50%¹³. Given our class imbalance (1:5.67 non-converters: converters), test sets’ classification performances were assessed using sensitivity and specificity, balanced accuracy (i.e., ($\mathit{\text{sensitivity}}+\mathit{\text{specificity}})/2$) and AUCROC. Calibration plots (Supplementary Methods 2) and decision curves (Supplementary Methods 3) were produced to assess model fit and clinical utility. Additional analyses were performed to assess the impact of participant selection on DSM performance (Supplementary Methods 1.6, Supplementary Methods 4) and to compare the importance of each data modality (Supplementary Methods 1.6, Supplementary Methods 5).

2.5. Binary classifiers

We benchmarked the DSM’s conversion classification performance against seven widely used classifiers: (i) random forest, (ii) logistic regression, (iii) support vector, (iv) decision tree, (v) gradient boosting, (vi) gaussian naive bayes, and (vii) k-nearest neighbors (with k = 5). The participants with pre-processed data were split into 20 folds identical to those provided to the DSM, and all were trained with a random state of 42. Models were implemented in the scikit learn library (sklearn)²⁶ using default parameters unless otherwise specified. The logistic regression model was trained for a maximum of 100,000 iterations. The support vector model was trained with a linear kernel and balanced class weights. The decision tree classifier had a maximum depth of 5. The gradient boosting classifier had a maximum depth of 7 and a learning rate of 0.01. The same performance metrics (sensitivity, specificity, balanced accuracy, and AUCROC) were averaged across the 20 test sets for each classifier.

2.6. Feature importance and phenotyping

Feature importance was evaluated using a permutation-based approach applied directly to the output of the DSM (Supplementary Methods 1.4). Phenotyping was used to investigate whether conversion rates were higher in groups of individuals with certain characteristics. Hazard ratios (HRs) compare the likelihood of conversion between two phenotypes (i.e., clinical, cognitive, or socio-demographic strata) over the time interval between baseline and t and are quantified by $HR=h_1\left(t\right)/h_0\left(t\right)$, where $h_0\left(t\right)$ is hazard rate in one phenotype and $h_0\left(t\right)$ is hazard rate in the comparison phenotype. The IntersectionalPhenotyper function from the auton_survival package was unsupervised, and recovered groups (i.e., phenotypes) of individuals over exhaustive combinations of categorical and numerical features. Numerical features were binned into ‘low’, ‘medium’, and ‘high’ terciles. Log-rank tests were implemented from the lifelines library to compare conversion rates between phenotypes, by testing for statistically significant differences in conversion rates between groups. In other words, the observed number of conversion events in each group was compared to the expected number of events based on the overall survival distribution.

3. Results

3.1. Participants

Of the total 1415 subjects, all of whom were CN at baseline, 15% converted within the study period (Table 1). Most subjects were female (68%), white (83%), and non-Hispanic (92%). Compared to converters, a significantly higher proportion of non-converters were female (69% and 60% respectively, z = 2.58, p = .01). There were no significant differences between converters and non-converters in the proportions of non-white (z = 1.56, p = .12) or Hispanic subjects (z = 0.79, p = .43). Converters had significantly fewer years of education (15.62) compared to non-converters (16.33; t = 3.26, p < .01), although most subjects had at least 12 years of education (98%).

Table 1.

Demographics including sex (assigned at birth), age, time in study, and years of education for all subjects, and converters and non-converters separately.

	all subjects N = 1415				converters N = 212				non-converters N = 1203
	mean	SD	min	max	mean	SD	min	max	mean	SD	min	max
female (%)	67.99				60.38				69.33
non-white (%)	16.67				20.28				15.96
Hispanic (%)	7.63				8.96				7.40
age	68.82	10.71	21.25	95.58	77.36	7.95	50.25	95.58	67.32	10.43	21.25	95.17
education	16.22	2.78	2.00	25.00	15.62	2.96	5.00	22.00	16.33	2.74	2.00	25.00
time
baseline to last visit	6.06	3.06	0.70	20.46	5.47	2.95	0.70	16.42	6.16	3.07	2.51	20.46
baseline to conversion					3.43	1.97	0.48	7.94
conversion interval					1.35	0.69	0.48	5.17

NOTE. values are in years unless otherwise specified.

At baseline, subjects were mostly mid-life-to-older adults (mean CA= 68.82 y, SD CA = 10.71 y); less than 2% of subjects were under 45 y (CA) at baseline, all of whom were non-converters. Converters were significantly older than non-converters, on average (77.36 y and 67.32 y respectively, t = 16.10, p < .01). The minimum time between baseline and final visits was 0.70 y (~8 months), and the maximum was 20.46 y. Non-converters remained in the study, on average, longer than converters (6.16 y and 5.47 y respectively, t = 3.12, p < .01). Converters received an MCI or AD diagnosis within an average of ~3.5 y, with the longest conversion being almost 8 y after baseline. The average interval between the penultimate visit before conversion and the visit at which CI was diagnosed (i.e., the conversion interval) was 1.35 y.

3.2. Model performance

The DSM achieved near-excellent predictive ability to discriminate between converters and non-converters, with a c-index of 0.88 (SD = 0.07), integrated Brier score of 0.07 (SD = 0.02), and AUCROC of 0.89 (SD = 0.06). Supplementary Table 3 displays performance metrics across each of the 20 folds. Figure 1 displays the survival probability plot for all converters and non-converters. In the test set, converters’ average survival probability dropped below 50% at 10 y follow-up, while non-converters’ average survival probability did not drop below 50% within the 22-y time frame. Hazard rates were lower in non-converters (ranging from 0.02 at 2 y to 0.03 at 22 y) compared to converters (0.09 at both 2 y and 22 y). The hazard ratios indicate that, at the 2-y mark, converters were at a 3.87 times higher rate of conversion than non-converters (p < .001). This multiplier diminished to 3.15 times the higher rate by the 22-y mark (p < .001). While the true conversion time was, on average, 3.43 y, the model’s sensitivity did not exceed 50% until the 10-year mark. The mean survival time predicted by the model was 8.81 y (SD = 6.95 y).

Figure 1.

Results of the deep survival machine predicting non-conversion (i.e., survival) probability over time. Panel (A) displays mean (and 95% confidence interval) survival probabilities for converters (red) and non-converters (blue) in the training and test sets. When survival probability decreases below 0.5, subjects are deemed a ‘converter’. Mean survival probabilities from the test set are also displayed as a histogram with 95% confidence intervals, from 2 years to 22 years in panel (B). Panel (C) depicts mean hazard rates for converters and non-converters over time, and hazard ratios (HR; i.e., converters’ mean hazard rate/non-converters mean hazard rate).

To calculate balanced accuracy, sensitivity, and specificity, converters were classified according to survival probability below 50% at 2-y increments over the study period (Supplementary Figure 1). At 2 y, no subjects had survival probability < 50%, resulting in 100% specificity and 0% sensitivity. Specificity decreased by ~2% to 3% with each 2-year increment in survival time. Sensitivity, on the other hand, increased more rapidly, reaching 75% by the 22-y mark. This resulted in a final balanced accuracy of 80%, where 148 of the 212 converters and 959 of the 1203 non-converters were correctly identified. Calibration plots reveal reasonably well-calibrated risk estimates at the evaluated time horizons, with a tendency for the predicted risks to be underestimated after approximately 16 years (Supplementary Methods 2 and Supplementary Figure 2). Decision curves show that the DSM is superior to both the “treat all” and “treat none” benchmarks for risk thresholds up to 50% (depending on the event horizon). The 10-year event horizon offers the most superior performance compared to benchmarks (Supplementary Methods 3 and Supplementary Figure 3).

The DSM had the highest balanced accuracy (75%) of all classifier models (Supplementary Table 4). Compared to the next best performing model (Gaussian naïve bayes, balanced accuracy of 71%), the DSM had higher specificity (80% compared to 70%) and only slightly lower sensitivity (70% compared to 72%). Because of the imbalance in the dataset (15% converters), all non-DSM classifiers except the Gaussian naïve Bayes model produced very high specificities but very low sensitivities. Thus, non-DSM models had balanced accuracies less than 70% and sensitivities of approximately 50% or less.

Sensitivity analyses revealed that the inclusion of reverters did not significantly alter the balanced accuracy of the model (Supplementary Methods 4, Supplementary Tables 8 – 10, and Supplementary Figure 4). Each ablated model exhibited significant depreciation, or no significant difference in balanced accuracy compared to the original model (Supplementary Figure 5, Supplementary Table 11).

3.3. Feature importance

Cognitive test scores, especially for logical memory (delayed), Benson figure drawing (delayed), logical memory (immediate), trail-making test B, and digits forward, were the most important features for predicting conversion (normalized importances = 1.00, 0.73, 0.72, 0.60, and 0.54 respectively, see Figure 2 and Supplementary Table 5). Subjects who scored lower on these tests were more likely to convert. Subjective cognitive decline was among the most important features for predicting conversion, with subjects who endorse significant memory changes over the last year being more likely to convert (normalized importance = 0.39). Of the 50 most important features, 38% were regional BVs, 32% were clinical and socio-demographic measures (e.g., cognitive test scores, self-reported cognitive status, age, geriatric depression score), 22% were regional BA saliency features, and 8% were measures relating to brain health (e.g., brain age, APOE status, amyloid positivity on PET, history of stroke).

Figure 2.

Top 50 most important clinical (purple), health (blue), and imaging (green) features for predicting conversion. Larger values represent higher predictive power and thus greater importance for conversion prediction than smaller values.

BA saliency and BV features’ importances (i.e., average minimum depths) were plotted on the cortex for comparison (Figure 3). Regions where BA saliency was important to conversion prediction included insular, temporal, and occipital gyri, the cingulate and sub-callosal gyri, and subcortical regions including the left parahippocampal gyrus and amygdala. Regions where BV was most important to conversion prediction were the hippocampi, superior frontal and temporal gyri, parietal and parieto-occipital regions, the cuneus, the pallidum, and ventricular volumes.

Figure 3.

Cortical plots of feature importance for brain age saliency (A, C, and E) and brain volume (B, D, and E). Views are canonical and the units of panels (E) and (F) are millimeters (voxel sizes are 1mm isotropic). Larger values (dark red) represent higher predictive power and thus greater importance for conversion prediction than smaller values (white and yellow).

3.4. Comparison of conversion risk across phenotypes

Log-rank tests were performed to compare hazard rates between sociodemographic, medical, clinical, cognitive, and biomarker phenotypes. Of 217 comparisons, 89 were significant after FDR correction (Table 2, see Supplementary Table 6 for all comparisons). The following hazard rates and ratios are derived 2 y after baseline, with hazard rates expressed as risk of conversion per year. CA and BA phenotyping produced the highest hazard ratios, with older subjects experiencing 9.58- and 9.90-times (respectively, p < .001 for both) greater risks of conversion 2 years after baseline than younger subjects. Subjects whose CA or BA was less than 64 y had a 0.7% risk of conversion per year, while those older than 73 y had a 6.8% risk in each year. A large difference between BA and CA (i.e., large positive AG) imparts a 4.16 times higher risk of conversion compared to a large negative AG (p < .001), and 2.17 times higher risk compared to typical brain aging (i.e., AG ~ 0; p < .001y).

Table 2.

Comparison of conversion occurrences and hazard rates and ratios across subjects of different socio-demographic, medical, clinical, cognitive, and biomarker strata.

variable	lower risk group > higher risk group	hazard rate lower risk group	hazard rate higher risk group	hazard ratio	χ2	p
socio-demographic strata
age	21 to 64 yrs > 73 to 96 yrs	0.007	0.068	9.582	154.128	<.001
	65 to 72 yrs > 73 to 96 yrs	0.020	0.068	3.363	64.078	<.001
	21 to 64 yrs > 65 to 72 yrs	0.007	0.020	2.849	26.816	<.001
brain age	0 to 63 yrs > 74 to 102 yrs	0.007	0.068	9.901	159.054	<.001
	64 to 73 yrs > 74 to 102 yrs	0.007	0.021	2.999	48.442	<.001
	0 to 63 yrs > 64 to 73 yrs	0.021	0.068	3.301	39.982	<.001
age gap	−31 to −3 yrs > 1 to 22 yrs	0.013	0.056	4.164	80.353	<.001
	−2 to 0 yrs > 1 to 22 yrs	0.026	0.056	2.169	40.528	<.001
	−31 to −3 yrs > −2 to 0 yrs	0.013	0.026	1.920	6.191	.013
sex	female > male	0.028	0.039	1.400	7.120	.008
residence	private > retirement community	0.029	0.098	3.408	32.097	<.001
	private > assisted living	0.029	0.152	5.274	66.585	<.001
	retirement community > assisted living	0.098	0.152	1.548	10.940	.001
	other > assisted living	0.034	0.152	4.494	24.357	<.001
marital status	married/de facto > widowed	0.028	0.060	2.144	35.044	<.001
	divorced/separated > widowed	0.030	0.060	2.036	17.747	<.001
	never married > widowed	0.024	0.060	2.526	9.565	.002
year of birth	1950 to 1997 > 1913 to 1940	0.008	0.067	8.832	98.308	<.001
	1950 to 1997 > 1941 to 1949	0.008	0.022	2.851	24.909	<.001
	1941 to 1949 > 1913 to 1940	0.022	0.067	3.098	35.080	<.001
race	white > Asian	0.032	0.051	1.630	17.521	<.001
	black > Asian	0.032	0.051	1.591	8.555	.003
medical
systolic blood pressure	78 to 122 mmHg > 139 to 220 mmHg	0.020	0.041	2.083	19.752	<.001
	78 to 122 mmHg > 123 to 138 mmHg	0.020	0.035	1.761	7.806	.005
	123 to 138 mmHg > 139 to 220 mmHg	0.035	0.041	1.183	6.043	.014
years of smoking cigarettes	0 yrs > 1 to 5 yrs	0.028	0.039	1.358	7.315	.007
history of heart attack/cardiac arrest	absent > remote/inactive	0.030	0.064	2.090	13.080	<.001
history of stroke	absent > remote/inactive	0.031	0.072	2.307	16.304	<.001
history of traumatic brain injury	no > yes	0.031	0.044	1.407	24.994	<.001
history of hypertension	absent > recent/active	0.024	0.044	1.849	21.210	<.001
	absent > remote/inactive	0.024	0.039	1.638	5.603	.018
hearing normal without hearing aid	yes > no	0.026	0.055	2.136	31.968	<.001
clinical
CDR: memory impairment	no impairment > questionable impairment	0.028	0.067	2.443	42.146	<.001
	no impairment > mild impairment	0.028	0.079	2.883	24.137	<.001
CDR: judgment and problem-solving	no impairment > questionable impairment	0.029	0.071	2.393	30.822	<.001
CDR: community affairs	no impairment > questionable impairment	0.031	0.101	3.255	20.563	<.001
global CDR	no impairment > questionable impairment	0.027	0.070	2.564	42.613	<.001
self-reported memory decline	no > yes	0.027	0.045	1.680	44.662	<.001
informant-reported memory decline	no > yes	0.029	0.054	1.885	44.826	<.001
decline in memory ability	no > yes	0.030	0.096	3.237	18.178	<.001
decline in visuospatial function	no > yes	0.031	0.128	4.068	10.094	.001
predominant symptom of cognitive decline	no impairment > memory impairment	0.030	0.099	3.343	21.332	<.001
	attention impairment > no impairment	0.014	0.030	2.200	6.697	0.01
mode of cognitive symptom onset	no impairment > gradual	0.030	0.079	2.644	25.340	<.001
mode of behavioral symptoms onset	no symptoms > gradual	0.031	0.046	1.498	9.457	.002
first symptom of motor decline	no symptoms > gait disorder	0.031	0.128	4.106	32.208	<.001
	tremor > gait disorder	0.044	0.128	2.883	6.983	.008
clinician-assessed cognitive status	better than normal > no opinion	0.012	0.038	3.119	12.914	<.001
	no opinion > one/two abnormal scores	0.038	0.062	1.619	7.611	.006
	no opinion > three+ abnormal scores	0.038	0.096	2.496	17.007	<.001
	better than normal > normal for age	0.012	0.029	2.333	18.578	<.001
	better than normal > one/two abnormal scores	0.012	0.062	5.050	69.766	<.001
	better than normal > three+ abnormal scores	0.012	0.096	7.785	89.280	<.001
	normal for age > one/two abnormal scores	0.029	0.062	2.165	46.839	<.001
	normal for age > three+ abnormal scores	0.029	0.096	3.337	36.172	<.001
	one/two abnormal scores > three+ abnormal scores	0.062	0.096	1.541	5.854	.016
cognitive
immediate story recall percent correct	61 to 96% > 7 to 47%	0.018	0.048	2.643	40.222	<.001
	48 to 60% > 7 to 47%	0.030	0.048	1.579	25.947	<.001
delayed story recall percent correct	56 to 96% > 0 to 42%	0.017	0.048	2.846	67.824	<.001
	56 to 96% > 43 to 55%	0.017	0.031	1.805	14.366	<.001
	43 to 55% > 0 to 42%	0.031	0.048	1.577	21.155	<.001
story recall delay time	20 to 57 mins > 7 to 16 mins	0.025	0.046	1.793	6.156	.013
digits forward percent correct	88 to 100% > 0 to 74%	0.022	0.042	1.906	18.519	<.001
	75 to 87% > 0 to 74%	0.032	0.042	1.316	10.341	.001
digits backward percent correct	66 to 100% > 25 to 58%	0.021	0.046	2.206	29.987	<.001
	66 to 100% > 59 to 65%	0.021	0.029	1.395	10.139	.001
	59 to 65% > 25 to 58%	0.029	0.046	1.581	6.014	.014
verbal fluency animals count	24 to 43 > 5 to 19	0.014	0.057	4.029	73.668	<.001
	24 to 43 > 20 to 23	0.014	0.023	1.591	9.398	.002
	20 to 23 > 5 to 19	0.023	0.057	2.531	38.405	<.001
verbal fluency vegetables count	18 to 32 > 3 to 13	0.012	0.057	4.721	71.941	<.001
	18 to 32 > 14 to 17	0.012	0.025	2.069	5.583	.018
	14 to 17 > 3 to 13	0.025	0.057	2.282	48.579	<.001
trail-making test A completion time	6 to 23 secs > 33 to 150 secs	0.011	0.058	5.273	116.078	<.001
	24 to 32 secs> 33 to 150 secs	0.027	0.058	2.136	15.654	<.001
	6 to 23 secs > 24 to 32 secs	0.011	0.027	2.469	54.542	<.001
trail-making test B completion time	20 to 55 secs > 81 to 328 secs	0.010	0.061	6.335	118.121	<.001
	56 to 80 secs > 81 to 328 secs	0.024	0.061	2.567	44.943	<.001
	20 to 55 secs > 56 to 80 secs	0.010	0.024	2.468	22.092	<.001
naming percent correct	98 to 100% > 25 to 93%	0.020	0.051	2.570	15.572	<.001
	94 to 97% > 25 to 93%	0.026	0.051	1.943	21.961	<.001
biomarker
elevated amyloid on PET	no > yes	0.026	0.061	2.336	10.369	0.001

NOTE. Table displays only statistically significant log-rank comparisons according to Chi squared (χ²) test statistics and related p-values (FDR corrected). See Supplementary Table 6 for complete results including non-significant comparisons, and for descriptive information for each variable. Direction ‘>’ indicates higher survival probability (i.e., ‘no impairment > mild impairment’ indicates a higher survival probability for subjects with ‘no impairment’ compared to ‘mild impairment’). For the purpose of phenotype comparisons, continuous data are divided into tertiles representing low (1^st to 33^rd percentile), medium (34^th to 67^th percentile), and high (67^th to 99^th percentile) scorers. The following continuous variables were not normally distributed and could not be converted into tertiles: neuropsychiatric inventory questionnaire total score, functional activities questionnaire total score, geriatric depression scale total score, CDR sum of boxes, Benson figure drawing immediate score, Benson figure drawing delayed score. Hazard rates are calculated at time = 2 years, significant hazard ratios are in bold (after FDR correction).

Of all phenotypes, subjects in assisted living had the highest 2-y hazard rate (15.2% risk in each year), followed by decline in visuospatial functioning (12.8% risk in each year), and gait disorder (12.8% risk in each year). Hazard ratios indicate that persons in assisted living have 5.27 times the risk of conversion compared to those living independently (p < .001). Widowhood had a higher 2-y hazard rate (6.0%) than marriage (2.8%), divorce (3.0%), or never being married (2.4%), with widows having 2.14 times higher risk of conversion than those who are married (p < .001). Having questionable or mild impairment in memory, judgment and problem solving, or community affairs domains also imparted higher hazard rates (ranging from 6.7% to 10.1%) than no impairment (2.7% to 3.1%). Subjects with three or more abnormal cognitive test scores had a 7.79 times higher risk of conversion than those who scored better-than-normal for their CA (p < .001), and 3.34 times higher risk than those whose scores were normal for their ages (p < .001). The trail-making test B had the highest hazard ratio of the cognitive tests, with subjects who take more than 81 s to complete the task having a risk of conversion 6.34 times higher than those whose completion time was less than 55 s (p < .001). Surprisingly, subjects with elevated amyloid on PET had only a 2.34 times greater risk of conversion compared to those without (p = .02): however, only 6% of subjects had PET data available. All hazard rates were obtained at t = 2 years; rates and ratios are somewhat inconsistent over different values of t (see Figure 1 and Supplementary Table 7).

4. Discussion

Our results are among the few that address the critical issue of predicting AD-related CI risk in unimpaired individuals, as most research in this area has focused on the conversion from MCI to AD (for reviews, see^27,28). The DSM predicted the probability of conversion to CI among CN adults, with classification accuracies that surpassed those of previously published survival models^15,16. By utilizing a DSM trained on extensive clinical and MRI data, we achieved an accuracy of 75% and an AUCROC of 0.89, exceeding the performance of previously published classifiers that identified between 68% and 71% of future CI converters^6,7.

4.1. Motivation

A significant challenge in identifying converters is that the early stages of cognitive decline are often subtle^19,29,30. The anatomic correlates of cognitive decline are also variable, which complicates the validation of a comprehensive model that captures all such features³¹. While predicting CI conversion among CN adults is more difficult than identifying MCI-to-AD converters, it remains crucial to forecast CN-to-CI conversion while there is still time for intervention. Models that predict MCI-to-AD conversion are less useful for early intervention compared to those predicting CN-to-CI conversion, as neurodegenerative processes are likely irreversible by the time symptoms appear. Within 10 years of an MCI diagnosis, 37.8% of individuals convert to AD, and 46.7% die¹⁴.

4.2. Accuracy of CI conversion prediction

The DSM estimated the risk of CI conversion among CN adults with greater concordance than previous DSMs, achieving an unweighted c-index of 0.75 and an IPCW c-index of 0.88. By comparison, the unweighted c-indexes of these previous models were 0.66¹⁵ and 0.73¹⁶. Unlike other DSMs, this DSM accommodates fully parametric estimation of survival times in the presence of censoring. It does so by utilizing high-dimensional data and nonlinear representations of input covariates without assuming constant proportional hazards over the subjects’ lifetimes.

Our accuracy of 0.75, AUCROC of 0.89, sensitivity of 0.70, and specificity of 0.80 were either better than or equivalent to those of other classifiers and traditional survival models. For instance, Lin et al⁷ used 15 clinical features to predict MCI conversion within 4 years, achieving an AUCROC of 76%, accuracy of 71%, sensitivity of 68%, and specificity of 72%. Pang et al⁶ employed 20 clinical features to predict conversion within 4 years, obtaining an AUCROC of 71%, accuracy of 72%, sensitivity of 71%, and specificity of 72%. Pankratz et al.⁹ utilized a Cox proportional hazards model to assess the risk of MCI among CN adults aged 70 to 89, based on demographics, clinical features, and APOE ε4 carrier status. Over a median follow-up period of 4.8 y, they achieved a c-index of 0.70.

Although the accuracy of predicting CN-to-CI conversion should be lower than that of predicting CN and/or MCI-to-AD conversion, our DSM outperformed those reported in previous studies^8,13. Mirabnahrazam et al.¹³ utilized a DSM³² trained to predict the probability of AD diagnosis in 401 CN and MCI subjects. By combining 21 MRI features with measures of cognition, demographics, and CSF, they achieved a c-index of 0.83. Similarly, Nakagawa et al.⁸ reported a c-index of 0.84 using a DSM to predict AD conversion among CN and MCI subjects. In contrast, we achieved an IPCW c-index of 0.88. The DSM thus offers the means to identify subjects at risk of CI before the onset of symptoms, achieving higher accuracy than models trained on individuals already exhibiting clinically significant cognitive decline. This is particularly useful given that current clinical trials of anti-amyloid therapeutics assume that treatment while a person is still CN significantly increases the likelihood of stalling the onset of cognitive symptoms³.

Decision curves (Supplementary Figure 3) suggest the DSM prioritizes capturing all possible converters and is less concerned with penalizing false positives, particularly when the event horizon – i.e., the classification cut-off – is 16 years. Thus, it should be used to stratify people who are more risk-averse, not those of minimal concern, for further investigation. It should also be noted that lower clinical utility was observed in the decision curves before the 6-year horizon (Supplementary Figure 3). Together, these results highlight that the best horizon-specific utility of the model is approximately 6 to 16 years in the future, with the 10-year event horizon offering superior performance compared to both the “treat all” and “treat none” benchmarks. This 10-year horizon reflects the point where the mean survival probability for converters drops below 50% (Figure 1). The classification threshold for reporting accuracy was set at the maximum time that a participant was retained in the study (i.e., 22 years). The 22-year horizon reflects the point where the model demonstrated the best balance of sensitivity and specificity, which is the optimal cut-off for classification – which should not be conflated with prediction of time-to-event. Accuracies across all event horizons are available in Supplementary Figure 1.

4.2. Neuroanatomical and clinical features of importance

We found that subjective memory decline was among the most important features for predicting conversion. Subjective cognitive impairment (SCI) is a significant risk factor for future CI diagnosis¹⁹. Adults with SCI are 4.5 times more likely to convert, and decline more rapidly, than adults without SCI³³. Within a 7-y window, 54.5% of adults who endorse SCI converted to CI within 7 y, whereas only 14.5% of adults without SCI converted³³. We observed that subjects who endorse memory decline are at 1.68 times higher risk of conversion than those who do not. Living in assisted living, having higher CDR scores, and scoring lower on cognitive tests also imparted significantly higher risks of conversion (see Table 2). Indeed, scores on cognitive tests, particularly memory-based tests, were the top five most important features for predicting conversion (Figure 2). The ablation studies (Supplementary Figure 5 and Supplementary Table 11) suggest that clinical data contributed the most to the DSM’s prediction accuracy, followed by a combination of clinical and saliency and/or volume.

We found that frontal, parietal, and temporal BVs were important for predicting conversion, in agreement with previous findings³⁴. Medial temporal volumes are often highlighted as effective early predictors of AD due to neurofibrillary tangle formation starting in the medial temporal lobe³⁵. However, amyloid plaques first affect posterior association cortices³⁶ and metabolic studies suggest that AD-related dysfunction is more frequent in parietal areas³⁷. Inclusion of MRI can enhance the accuracy of conversion predictions by leveraging pre-symptomatic CI-related atrophy³⁸. Our results agree with those of previous reports indicating that structural brain changes, particularly in the hippocampus and medial temporal lobe, are apparent in CN adults preceding CI conversion^1,17,18. Temporal and hippocampal volumes and BA saliencies were important for predicting conversion, in line with reports that AD-related neurodegeneration begins in these areas^19,31,35,36.

Inclusion of BA measures can increase the accuracy of predicting CI¹⁹. We found that subjects whose brains appeared older than expected (i.e., AG > 0 y) had a 4.16 times greater risk of conversion compared to those whose brains appeared younger than expected (i.e., AG < 0 y). Having AG > 0 y also imparted a 2.17 times higher risk of conversion compared to normal brain aging (AG = 0 y). Our results highlight that BA saliency captures neuroanatomical abnormalities in converters across a range of brain regions beyond temporal or parietal cortices. BA saliency of insular, cingulate, and subcortical regions were important for conversion prediction.

4.3. Limitations and future directions

Conversion cannot be predicted perfectly based on baseline observation data alone because risk changes over time. Baseline clinical, cognitive, and neuroimaging measures provide incomplete information needed to predict conversion, since a portion of the risk lies in future environmental conditions. Longitudinal survival modelling with three timepoints has improved prediction of AD onset among MCI patients from 75.6% to 97.1%³⁹, and is a promising direction for future attempts to predict CI among CN subjects. Similarly, modelling ‘reversion’ is not a trivial issue, as this subgroup likely contains diagnostically ambiguous or noise-prone cases (Supplementary Figure 4), but their exclusion may introduce survivorship bias, potentially inflating model accuracies. Predicting risk in this subgroup is important, but requires a sufficiently large, repeated measures sample to train a recurrent survival model to track changes in survival risk across multiple time points. Also, we did not include mortality as a competing risk because this data was not available for all participants in our dataset. This may bias the estimated conversion probabilities when deaths are incorrectly treated as non-informative censoring. Future models should be suitably refined using mortality outcomes to avoid the overestimation of conversion and predict dementia risk in more realistic, complex, multi-morbidity scenarios.

Racial/ethnic minorities, including Hispanic individuals of any ancestry, are under-represented in AD research⁴⁰ but are more likely to be diagnosed with CI⁴¹. Only 23% of our participants were non-White and/or Hispanic. NACC participants are also skewed towards the higher education strata, and thus may have higher cognitive reserve than the general population⁴². Future studies should examine the generalizability of our findings to diverse populations and other cohorts, such as the Alzheimer’s Disease Neuroimaging Initiative (http://adni.loni.usc.edu/) and the Open Access Series of Imaging Studies (https://sites.wustl.edu/oasisbrains/). Validating the DSM on these external datasets will also help avoid overfitting by increasing data size and diversity; however, this requires data harmonization, particularly of the clinical data, that fell beyond the scope of this study.

Of note, the DSM did not accurately estimate actual time until conversion. The predicted survival time for converters was overestimated by ~ 5 y, likely due to the higher proportion of non-converters within the sample. Our findings endorse the use of the DSM for classifying converters versus non-converters and for estimating conversion rate based on subject phenotypes. However, the DSM is not suitable for estimating the time remaining before CI symptom onset. Partly, this is because time to conversion was not measured precisely – its recorded value depends on the timing of participants’ examinations. This issue can be addressed using interval censoring, where conversion is known to lie within the interval between visits, into the DSM algorithm.

Overfitting, due to relatively small sample size and large number of features, was a concern that was addressed by 1) using 20-fold cross validation to obtain a more robust assessment of the model’s performance on unseen data, 2) early stopping to prevent the model from learning noise in the training data, 3) LASSO regularization to stop the model parameters becoming too large, and 4) carefully tuning the model’s hyperparameters and adjusting the learning rate to find a balance between model complexity and data fit. Overall, the model produced reasonably well-calibrated risks at the evaluated time horizons. Future iterations of the model could investigate scaling or isotonic regression to bring the probabilities in line with actual outcomes.

Finally, we acknowledge that multi-site acquisition differences may influence prediction accuracy. To maximize sample size, we used neuroimages from NACC’s mixed, non-harmonized MRI protocol. FreeSurfer’s volumetric outputs are robust to multi-site acquisition differences, demonstrating high test-retest reproducibility across scanners and acquisition protocols (mean ICCs>0.97)⁴³. Predicted brain age is also highly reliable across scanners (ICC = 0.92)⁴⁴, and while the cross-site reliability of brain-age saliency maps remains to be empirically assessed, saliency estimation tends to be reliable⁴⁵.

4.4. Conclusion

Early identification of future CI converters is crucial for timely intervention and for enriching clinical trial samples, as symptomatic AD is characterized by advanced, irreversible neurodegeneration. While most studies in this area focus on identifying progressive MCI cases among symptomatic subjects, we forecast the future risk of CI among unimpaired CN adults. The DSM predicts conversion to MCI or AD among CN adults with 75% accuracy, an AUCROC of 0.89, and a c-index of 0.88, outperforming previously published models. This work shows that advances in ML and survival modeling can be utilized to enhance the lives of aging individuals by using clinical and neuroimaging data to identify preclinical AD, particularly in risk-averse individuals.

Diversity, equity, and inclusion statement:

Diversity, equity, and inclusion (DEI) was addressed in the study design by examining the association between race/ethnicity and diagnosis of cognitive impairment. The ancestry profile of the cohort studied here is not representative of the US population, and this limitation is acknowledged in the study. ‘Female’ and ‘male’ refer to participants sex assigned at birth. All participants in our analysis selected either ‘male’ or ‘female’. However, individuals were not excluded if they chose not to answer this question.

Data availability:

Qualified researchers may obtain access to all de-identified imaging data and demographic and cognitive used for this study at https://naccdata.org/.