Predicting Time-to-conversion for Dementia of Alzheimer's Type using Multi-modal Deep Survival Analysis
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: We're looking at the paper: "Predicting Time-to-conversion for Dementia of Alzheimer's Type using Multi-modal Deep Survival Analysis".
Jane: Here's the gist of the paper. Dementia of Alzheimer’s Type (DAT) is a complex disorder influenced by numerous factors, but it is unclear how each factor contributes to disease progression Our findings showed that genetic features contributed the least to survival analysis, while CDC features contributed the most The study used 401 subjects with 63 features from MRI, genetic, and CDC data modalities in the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database.
Tom: Study Design and Data Modalities The researchers utilized a multi-modal deep learning-based survival analysis model that extends the classic Cox regression model to predict time-to-conversion to DAT.
Jane: The dataset comprised 401 subjects with 63 features drawn from MRI, genetic, and CDC data modalities in the ADNI database.
Tom: Data stratification involved dividing subjects into five subgroups based on clinical diagnosis at baseline, which were then divided into two categories based on whether they received a clinical diagnosis of DAT in a future time point.
Paper summary: Tom: So we’ve covered the overview, but let’s really nail down what this paper is proposing. Essentially, it’s about predicting when someone will convert to Dementia of Alzheimer's Type by using a multi-modal deep survival analysis model.
Jane: Right. The core thesis is that combining different data modalities—MRI scans, genetic information, and clinical data like CDC features—can give us a better prediction for the time it takes for someone to develop DAT.
Lu: The paper claims that while many factors influence Alzheimer's Type progression, the exact contribution of each factor to this progression is still unknown. They are using this model to investigate that contribution by looking at how different data types work together in predicting DAT risk.
Meng: So they aren't just saying "MRI is good" or "Genetics is good"; they’re trying to map out the actual influence of each piece of information within the prediction framework.
Lalam: They used a dataset from the Alzheimer's Disease Neuroimaging Initiative, ADNI, with four hundred one subjects and sixty three features pulled from MRI, genetic data, and CDC data modalities. This is their input pool for testing this multi-modal approach.
Tom: And the key result they highlight is that CDC features contributed the most to the model performance in their survival analysis. That’s a major finding when you’re trying to prioritize which inputs are most influential.
Jane: It matters because it suggests that clinical data, specifically those cognitive test features from CDC, have a stronger signal for predicting DAT progression than genetic or structural imaging features alone in this setup.
Lu: They also found that cognitive performance tests were very important, with all ten COG features declared to have a positive effect on the outcome when using the CDC feature set. This points to the value of those specific cognitive measures.
Meng: So it’s not just about having lots of data; it's about knowing which subset is actually driving the prediction, and this paper gives us that breakdown by modality.
Lalam: They also found that genetic factors and CSF factors were less useful when they were looked at in single data types, emphasizing the need for integration to see their true impact.
Tom: And they showed that using current clinical procedures with cognitive test results can actually outperform survival analysis results generated when only using costly genetic or CSF data. That’s a big practical implication for how clinicians might approach initial risk assessment.
Jane: It means sticking to what's already being done clinically—getting those cognitive tests first—might give you a better initial prediction than relying solely on expensive biological information right off the bat.
Lu: They also showed that combining MRI and CDC features yielded the best overall performance, with just the CDC features being very close behind. That tells us about the optimal feature set for this model structure.
Meng: So for a production system, it seems like MRI and CDC is a solid starting point based on these results because of how well they performed against other combinations.
Lalam: The paper also showed that genetic data has the potential to detect risk in currently normal subjects, which opens up possibilities beyond just predicting progression in already sick patients.
Tom: So, to wrap this part up, the main claim is that multi-modal deep survival analysis can predict time-to-conversion for DAT, and it helps us understand which specific data features are most predictive.
Jane: And it matters because it shows that we can move toward models that use clinical data effectively to predict progression earlier than before.
Conclusion: Tom: Thinking about this whole thing, "Predicting Time-to-conversion for Dementia of Alzheimer's Type using Multi-modal Deep Survival Analysis" and the authors, Mirabnahrazam a et al., Da Ma a et al., Beaulac a et al., and the others.
Jane: They’ve done this complex work by building this deep learning survival model to see how different data types interact. The implications are that we can start using these tools to get earlier risk assessments for DAT, even if we don't have a cure yet.
Lu: The key implication is moving away from single-data-type reliance toward integrated systems where the prediction relies on combining clinical, genetic, and imaging data for better accuracy.
Meng: From an engineering standpoint, it means that when we build these tools, we need to ensure the model is robust enough to handle that mix of inputs without breaking at scale.
Lalam: It suggests that cognitive tests are a good starting point because they're fast and cheap, which makes them a pragmatic input for early risk prediction workflows.
Tom: So in simple terms, this paper shows us that we can use sophisticated AI to get an earlier warning signal for DAT by looking at how different data streams feed into the prediction.
Jane: And it matters because it gives us a concrete direction on which data sources are most likely to provide that
Ghazal Mirabnahrazam, Da Ma, Cédric Beaulac, Sieun Lee, Karteek Popuri, Hyunwoo Lee, иgiguo Cao, Нames E Galvin, 㹮 Lei Wang, Mirza Faisal Beg
School of Engineering, Simon Fraser University · School of Medicine, Wake Forest University · Department of Mathematics and Statistics, University of Victoria · Mental Health & Clinical Neurosciences, School of Medicine, University of Nottingham · Department of Computer Science, Memorial University of Newfoundland · Division of Neurology, Department of Medicine, University of British Columbia · Department of Statistics and Actuarial Science, Simon Fraser University · Comprehensive Center for Brain Health, Department of Neurology, University of Miami Miller School of Medicine · НPsychiatry and Behavioral Health, Ohio State University Wexner Medical Center
cs.LG, cs.CV, q-bio.NC
Submitted: 2022-05-02
Updated: 2022-05-02
Journal ref: Neurobiology of Aging, 121, (2023), 139-156
DOI: 10.1016/j.neurobiolaging.2022.10.005
Code: https://github.com/havakv/pycox
Project page: http://www2.ensc.sfu.ca
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 77/100
The gist: Dementia of Alzheimer’s Type (DAT) is a complex disorder influenced by numerous factors, but it is unclear how each factor contributes to disease progression Our findings showed that genetic…
Key concepts
- Deep Learning-based Survival Analysis Model
- This is a sophisticated statistical method that uses a neural network to predict the time until an event occurs, like converting to DAT. It goes beyond traditional methods by learning complex patterns from multiple data types (MRI, genetics) simultaneously to estimate survival rates over time.
- Feature Importance
- This analysis determined which input variables—like MRI scans or genetic markers—had the biggest impact on the model's predictions. The study found that features from clinical cognitive tests were the most influential in accurately predicting DAT conversion.
- Time-to-Conversion (TTC)
- This is a crucial prediction: estimating exactly how long it will take for a patient to develop Dementia of Alzheimer's Type. The model helped identify risks early, with some subjects showing predicted risk earlier than the actual event occurred.
Terminology
Summary
Dementia of Alzheimer’s Type (DAT) is a complex disorder influenced by numerous factors, but it is unclear how each factor contributes to disease progression Our findings showed that genetic features contributed the least to survival analysis, while CDC features contributed the most The study used 401 subjects with 63 features from MRI, genetic, and CDC data modalities in the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. p. 0 Predicting Time-to-conversion for Dementia of Alzheimer’s
Type using Multi-modal Deep Survival Analysis
Ghazal Mirabnahrazam a,1, Da Ma a,b, Cédric Beaulac a,c, Sieun Lee a,d, Karteek Popuri a,e, Hyunwoo Lee f
, Jiguo
Cao g
, James E Galvin h
, Lei Wang i
, Mirza Faisal Beg a,*, and the Alzheimer’s Disease Neuroimaging Initiative 2
a School of Engineering, Simon Fraser University, Burnaby, BC, Canada
b School of Medicine, Wake Forest University, Winston-Salem, NC, USA
c Department of Mathematics and Statistics, University of Victoria, Victoria, BC, Canada
d Mental Health & Clinical Neurosciences, School of Medicine, University of Nottingham, Nottingham, United Kingdom
e Department of Computer Science, Memorial University of Newfoundland, St. John’s, NL, Canada
f Division of Neurology, Department of Medicine, University of British Columbia, Vancouver, BC, Canada
g Department of Statistics and Actuarial Science, Simon Fraser University, Burnaby, BC, Canada
h Comprehensive Center for Brain Health, Department of Neurology, University of Miami Miller School of Medicine, Miami, FL, USA
iPsychiatry and Behavioral Health, Ohio State University Wexner Medical Center, Columbus, OH, USA
Abstract
Dementia of Alzheimer’s Type (DAT) is a complex disorder influenced by numerous factors,
but it is unclear how each factor contributes to disease progression.Where the paper says it
Study Design and Data Modalities
The researchers utilized a multi-modal deep learning-based survival analysis model that extends the classic Cox regression model to predict time-to-conversion to DAT. The dataset comprised 401 subjects with 63 features drawn from MRI, genetic, and CDC data modalities in the ADNI database. p. 0 An in-depth examination of these
factors may yield an accurate estimate of time-to-conversion to DAT for patients at various disease
stages. We used 401 subjects with 63 features from MRI, genetic, and CDC (Cognitive tests,
Demographic, and CSF) data modalities in the Alzheimer’s Disease Neuroimaging Initiative (ADNI)
database. Subjects from the stable(s)NC,
unstable(u)NC, progressive(p)NC, stable(s)MCI, and progressive(p)MCI stratified subgroups were
included in this study. These stratified subgroups were divided into two categories based on whether they
received a clinical diagnosis of DAT in a future time point or not.Where the paper says it
Where the paper says it
Model Training and Evaluation
The base neural network employed was a multilayer perceptron (MLP) with the same number of nodes in each hidden layer, rectified linear unit (ReLU) activation function, and batch normalization between layers. The output of the network is a single node that estimates the hazard rate in the Cox model. p. 7 2018) which extends the classic Cox proportional hazard model to predict,
analyze, and compare the time-to-conversion to DAT. Cox regression model (Cox 1972)
provides a semi-parametric specification of the hazard rate:
ℎ(𝑡|𝑥) = ℎ0
(𝑡)𝑒𝑥𝑝[𝑔(𝑥)], 𝑔(𝑥) = 𝛽
𝑇𝑥, (1)
where ℎ0
(𝑡) is a non-parametric baseline hazard, and exp[𝑔(𝑥)] is the hazard ratio or risk score. Therefore, here we use the time-dependent concordance index (Ctd-index) (Antolini, Boracchi,
and Biganzoli 2005), which estimates the probability that observations 𝑖 and 𝑗 are concordant
provided that they are comparable:
𝐶
𝑡𝑑 = 𝑃{ 𝑆̂(𝑇𝑖
| 𝑥𝑖
) < 𝑆̂(𝑇𝑗
| 𝑥𝑗) | 𝑇𝑖 < 𝑇𝑗
, 𝐷𝑖 = 1}. 2 Brier score
The Brier score (BS) (Brier 1950) is used to evaluate the accuracy of a predicted survival
function; it represents the average squared distances between the observed survival status, 𝑦𝑖 ∈
{0,1}, and the predicted survival probability, 𝑝̂𝑖 and is always a number between 0 and 1, with 0
being the best possible value. 𝐻(𝑡|𝑥) = ∫ ℎ(𝑠|𝑥) 𝑑𝑠 =
𝑡
0
∫ ℎ0
(𝑠) 𝑒𝑥𝑝[𝑔(𝑥)] 𝑑𝑠,
𝑡
0
(4)
and in practice can be estimated by:
𝐻̂(𝑡|𝑥) = ∑ ∆𝐻̂0
(𝑇𝑖 𝑇
) 𝑒𝑥𝑝[𝑔̂(𝑥)]
𝑖≤𝑡
, (5)
where ∆𝐻̂0
(𝑇𝑖
) is an increment of the Breslow estimate (Breslow 1972):
∆𝐻̂0
(𝑇𝑖
) =
𝐷𝑖
∑𝑗𝜖𝑅 𝑒𝑥𝑝[ ̂𝑔(𝑥𝑗) ]
𝑖
, (6)
and 𝑔̂(𝑥) is the estimate of 𝑔(𝑥) obtained from the neural network. In order to evaluate the performance of the above model, we have trained our data using
the Cox model (Cox 1972) as benchmark, as well as four other best-performing deep learningbased models in the literature and reported the results in section D of Supplementary Material.Where the paper says it
Where the paper says it
Where the paper says it
Feature Importance Findings
The analysis of feature importance revealed that CDC features contributed the most to the model performance. Specifically, cognitive test features were found to play an important role in the analysis, with all ten COG features declared to be have a positive effect on the outcome using the CDC feature set. p. 4 3 Feature pre-processing for the Multi-modal input data
We used all of the available features from the Cognitive tests (10 features), Demographic (4
features), and CSF (7 features) categories in the ADNI database, and created a new set of 21 features
called CDC. Then, in order to perform a fair comparison between feature sets, we selected the top 21
most important features from MRI and genetic data modalities. 3 Feature pre-processing for the Multi-modal input data
We used all of the available features from the Cognitive tests (10 features), Demographic (4
features), and CSF (7 features) categories in the ADNI database, and created a new set of 21 features
called CDC. Then, in order to perform a fair comparison between feature sets, we selected the top 21
most important features from MRI and genetic data modalities.Where the paper says it
Where the paper says it
Clinical Application and Time-to-Conversion
The study demonstrated that using the current clinical procedure, which includes gathering cognitive test results, can outperform survival analysis results produced using costly genetic or CSF data. The model was used to estimate the survival rate for each subject over a 10-year period across seven different feature sets. p. 7 To construct a non-linear version of the Cox model, the linear predictor 𝑔(𝑥) = 𝛽
𝑇𝑥 in
the relative risk function above is replaced by a 𝑔(𝑥) parametrized by a neural network. The
predictions are obtained by estimating the survival function, 𝑆̂(𝑡|𝑥) = exp[−𝐻̂(𝑡|𝑥)], where 𝐻̂(𝑡|𝑥)
is the cumulative hazard function, which is commonly used for specifying different survival
models, and is defined as: Progressive (uncensored): This group includes subjects from the pNC and pMCI groups who
received a DAT diagnosis after their initial visit during the study window. The predict labels for each subject in the study include: a) an event indicator of 0 or 1 indicating
whether the subject was progressive (1) or non-progressive (0), and b) a duration indicating the time
between the first visit and the time when the DAT diagnosis was confirmed for progressive subjects, and
the time between the first and last visit for non-progressive subjects.Where the paper says it
Where the paper says it
Conclusion on Modality Utility
The combination of MRI and CDC features (brown) had the best overall performance, closely followed by the results obtained using only CDC features. The genetic data (GEN) has the potential to detect the risk of developing DAT in the future for currently normal subjects, while other modalities have a lower predictive power. p. 3 2020; Lu and Colliot 2021). To
the best of our knowledge, this is the first study that performs a comprehensive analysis on the prediction
of the time to conversion to DAT for subjects in various stages of the disease using multi-modal data and
compares the predictive power of each data modality and the effect of each modality on the disease
diagnosis and progression. Genetic data including Single Nucleotide Polymorphism (SNP) + APOE information (GEN);
3. Cognitive tests (COG) such as the Mini Mental State Examination (MMSE);
4. 2020; Lu and Colliot 2021). To
the best of our knowledge, this is the first study that performs a comprehensive analysis on the prediction
of the time to conversion to DAT for subjects in various stages of the disease using multi-modal data and
compares the predictive power of each data modality and the effect of each modality on the disease
diagnosis and progression. The stratified subgroups were divided into two groups based on
whether or not they received a clinical diagnosis of DAT in a future time point. The
non-progressive group includes subjects from the sNC, uNC, and sMCI stratified
groups who did not receive a clinical diagnosis of DAT during the study period, and
the progressive group includes subjects from the pNC and pMCI groups who
received a DAT diagnosis after their initial visit during the study window.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Limitations and Future Directions
Our study has some limitations, including being limited by the sample size and characteristics of the group of subjects selected from the ADNI database. The standard deviation appears to be high for all cases in Figures 4 and 5, which may be an indication of the Permutation Importance method's high sensitivity to the input data. An approach to address this limitation would be to increase the sample size by using large-scale AD related databases such as UK Biobank database. The authors suggest that cognitive tests can be gathered first to determine the state of a patient’s health. p. 3 Genetic data including Single Nucleotide Polymorphism (SNP) + APOE information (GEN);
3. Cognitive tests (COG) such as the Mini Mental State Examination (MMSE);
4. John’s, NL, Canada
f Division of Neurology, Department of Medicine, University of British Columbia, Vancouver, BC, Canada
g Department of Statistics and Actuarial Science, Simon Fraser University, Burnaby, BC, Canada
h Comprehensive Center for Brain Health, Department of Neurology, University of Miami Miller School of Medicine, Miami, FL, USA
iPsychiatry and Behavioral Health, Ohio State University Wexner Medical Center, Columbus, OH, USA
Abstract
Dementia of Alzheimer’s Type (DAT) is a complex disorder influenced by numerous factors,
but it is unclear how each factor contributes to disease progression. An in-depth examination of these
factors may yield an accurate estimate of time-to-conversion to DAT for patients at various disease
stages. ca (Mirza Faisal
Beg)
1 First Author
2 Data used in preparation of this article were obtained from the Alzheimer’s disease Neuroimaging Initiative (ADNI) database
(http://adni. As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data
but did not participate in analysis or writing of this report. ca (Mirza Faisal
Beg)
1 First Author
2 Data used in preparation of this article were obtained from the Alzheimer’s disease Neuroimaging Initiative (ADNI) database
(http://adni. As such, the investigators within the ADNI contributed to the design and implementation of ADNI and/or provided data
but did not participate in analysis or writing of this report.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Model Comparison
The traditional Cox proportional hazard (CoxPH) model performed very similarly to our main model (DeepSurv; Katzman et al. 2018). DeepSurv outperformed CoxPH in 4 of 7 cases, but the differences were not statistically significant. p. 1 1 Introduction
Alzheimer's Disease (AD) or Dementia of Alzheimer’s Type (DAT) is a progressive
neurodegenerative condition characterized by psychiatric, cognitive and structural deteriorations that
accounts for 60% to 80% of all dementia cases. One out of every three seniors dies with Alzheimer’s
disease or another type of dementia, accounting for more deaths than breast and prostate cancer
combined (Alzheimer’s Association 2021). The models have been shown to be more successful at accurately
estimating the underlying relationship between covariates and the event of interest in complex
problems than the traditional models. We have utilized a deep learning based survival analysis model called DeepSurv
(Katzman et al. The models have been shown to be more successful at accurately
estimating the underlying relationship between covariates and the event of interest in complex
problems than the traditional models. We have utilized a deep learning based survival analysis model called DeepSurv
(Katzman et al. The models have been shown to be more successful at accurately
estimating the underlying relationship between covariates and the event of interest in complex
problems than the traditional models. We have utilized a deep learning based survival analysis model called DeepSurv
(Katzman et al.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Data Pre-processing Impact
The main method performed the best for CDC and MRI+CDC feature sets. The difference was statistically significant in 2 out of 4 cases between the main and option 1 methods, and 3 out of 4 cases between the main and option 2 methods. p. 1 1 Introduction
Alzheimer's Disease (AD) or Dementia of Alzheimer’s Type (DAT) is a progressive
neurodegenerative condition characterized by psychiatric, cognitive and structural deteriorations that
accounts for 60% to 80% of all dementia cases. One out of every three seniors dies with Alzheimer’s
disease or another type of dementia, accounting for more deaths than breast and prostate cancer
combined (Alzheimer’s Association 2021). 3 Feature pre-processing for the Multi-modal input data
We used all of the available features from the Cognitive tests (10 features), Demographic (4
features), and CSF (7 features) categories in the ADNI database, and created a new set of 21 features
called CDC. Then, in order to perform a fair comparison between feature sets, we selected the top 21
most important features from MRI and genetic data modalities. Therefore, here we use the time-dependent concordance index (Ctd-index) (Antolini, Boracchi,
and Biganzoli 2005), which estimates the probability that observations 𝑖 and 𝑗 are concordant
provided that they are comparable:
𝐶
𝑡𝑑 = 𝑃{ 𝑆̂(𝑇𝑖
| 𝑥𝑖
) < 𝑆̂(𝑇𝑗
| 𝑥𝑗) | 𝑇𝑖 < 𝑇𝑗
, 𝐷𝑖 = 1}. 2 Brier score
The Brier score (BS) (Brier 1950) is used to evaluate the accuracy of a predicted survival
function; it represents the average squared distances between the observed survival status, 𝑦𝑖 ∈
{0,1}, and the predicted survival probability, 𝑝̂𝑖 and is always a number between 0 and 1, with 0
being the best possible value.Where the paper says it
Where the paper says it
Where the paper says it
Survival Analysis Benefits
Using survival analysis techniques to estimate time-to-conversion to Alzheimer’s disease has the following advantages over other methods such as regression. First, there is no need to exclude subjects who did not develop DAT during the study period when using survival analysis which allows us to benefit from data collected from all study participants and ensure we do not accidentally bias the study by excluding these subjects. p. 3 2020; Lu and Colliot 2021). To
the best of our knowledge, this is the first study that performs a comprehensive analysis on the prediction
of the time to conversion to DAT for subjects in various stages of the disease using multi-modal data and
compares the predictive power of each data modality and the effect of each modality on the disease
diagnosis and progression. In current clinical practice, a DAT diagnosis cannot be made until
the patient exhibits clear signs of cognitive decline, which can be attributed in part to the multifactorial
nature of DAT. Methods for predicting the probability of a patient developing Alzheimer's disease as a
function of time, known as survival analysis, are important tools in helping understand the characteristics
of DAT. Therefore, here we use the time-dependent concordance index (Ctd-index) (Antolini, Boracchi,
and Biganzoli 2005), which estimates the probability that observations 𝑖 and 𝑗 are concordant
provided that they are comparable:
𝐶
𝑡𝑑 = 𝑃{ 𝑆̂(𝑇𝑖
| 𝑥𝑖
) < 𝑆̂(𝑇𝑗
| 𝑥𝑗) | 𝑇𝑖 < 𝑇𝑗
, 𝐷𝑖 = 1}. 2 Brier score
The Brier score (BS) (Brier 1950) is used to evaluate the accuracy of a predicted survival
function; it represents the average squared distances between the observed survival status, 𝑦𝑖 ∈
{0,1}, and the predicted survival probability, 𝑝̂𝑖 and is always a number between 0 and 1, with 0
being the best possible value.Where the paper says it
Where the paper says it
Where the paper says it
Progressive vs Non-Progressive Survival
A good survival model should predict a high survival rate for non-progressive subjects, whereas the progressive subjects should be associated with a low survival rate. Over a two-year period (top left graph), more than 85% of non-progressive subjects had a chance of survival greater than 50%. p. 5 The stratified subgroups were divided into two groups based on
whether or not they received a clinical diagnosis of DAT in a future time point. The
non-progressive group includes subjects from the sNC, uNC, and sMCI stratified
groups who did not receive a clinical diagnosis of DAT during the study period, and
the progressive group includes subjects from the pNC and pMCI groups who
received a DAT diagnosis after their initial visit during the study window. We used a deep learning-based survival analysis model that extends the classic Cox
regression model to predict time-to-conversion to DAT. Our findings showed that genetic features
contributed the least to survival analysis, while CDC features contributed the most. 1982) is arguably the most widely used
metric for global assessment of prognostic models in survival analysis. The C-index estimates
the likelihood that the predicted survival times of a random pair of individuals will have the same
ordering as their true survival times among all pairs of subjects that can be ordered.Where the paper says it
Where the paper says it
Where the paper says it
Time Difference Accuracy
The true and predicted event times were very close to each other for 40 subjects (26.7%), with a difference of close to zero. In addition, more than half of the progressive subjects (80 out of 150, or 53. p. 3 In addition, ADNI aims to provide researchers with
the opportunity to combine genetics with imaging and clinical data to help investigate mechanisms of the
disease. 2 Data selection and stratification
A total of 401 subjects from the first phase of ADNI (ADNI1; Mueller et al. Predicting Time-to-conversion for Dementia of Alzheimer’s
Type using Multi-modal Deep Survival Analysis
Ghazal Mirabnahrazam a,1, Da Ma a,b, Cédric Beaulac a,c, Sieun Lee a,d, Karteek Popuri a,e, Hyunwoo Lee f
, Jiguo
Cao g
, James E Galvin h
, Lei Wang i
, Mirza Faisal Beg a,*, and the Alzheimer’s Disease Neuroimaging Initiative 2
a School of Engineering, Simon Fraser University, Burnaby, BC, Canada
b School of Medicine, Wake Forest University, Winston-Salem, NC, USA
c Department of Mathematics and Statistics, University of Victoria, Victoria, BC, Canada
d Mental Health & Clinical Neurosciences, School of Medicine, University of Nottingham, Nottingham, United Kingdom
e Department of Computer Science, Memorial University of Newfoundland, St. John’s, NL, Canada
f Division of Neurology, Department of Medicine, University of British Columbia, Vancouver, BC, Canada
g Department of Statistics and Actuarial Science, Simon Fraser University, Burnaby, BC, Canada
h Comprehensive Center for Brain Health, Department of Neurology, University of Miami Miller School of Medicine, Miami, FL, USA
iPsychiatry and Behavioral Health, Ohio State University Wexner Medical Center, Columbus, OH, USA
Abstract
Dementia of Alzheimer’s Type (DAT) is a complex disorder influenced by numerous factors,
but it is unclear how each factor contributes to disease progression. Progressive (uncensored): This group includes subjects from the pNC and pMCI groups who
received a DAT diagnosis after their initial visit during the study window. The predict labels for each subject in the study include: a) an event indicator of 0 or 1 indicating
whether the subject was progressive (1) or non-progressive (0), and b) a duration indicating the time
between the first visit and the time when the DAT diagnosis was confirmed for progressive subjects, and
the time between the first and last visit for non-progressive subjects. The hazard ratio is the parametric part of the model which consists of a linear predictor
𝑔(𝑥) = 𝛽
𝑇𝑥. The Cox partial likelihood, with Breslow’s method for handling tied event times
(Breslow 1972), is given by:
𝐿𝑐𝑜𝑥 = ∏ (
𝑒𝑥𝑝[𝑔(𝑥𝑖
)]
∑𝑗𝜖𝑅 𝑒𝑥𝑝[ 𝑔(𝑥𝑗) ]
𝑖
)
𝐷𝑖
𝑖
, (2)
and the negative partial log-likelihood can then be used as a loss function:
𝑙𝑜𝑠𝑠 = ∑ 𝐷𝑖
𝑙𝑜𝑔 ( ∑ 𝑒𝑥𝑝[ 𝑔(𝑥𝑗) − 𝑔(𝑥𝑖
) ] 𝑗𝜖𝑅𝑖
𝑖 ), (3)
where 𝑖 denotes an individual, 𝐷𝑖
is an indicator labelling the observed event time as an
progressive (uncensored; 𝐷𝑖 = 1) or non-progressive (right-censored; 𝐷𝑖 = 0) observation, and
𝑅𝑖
is the set of all individuals at risk at time 𝑇𝑖
(not censored and have not experienced the event
before time 𝑇𝑖
). Keywords: Alzheimer’s Disease, Deep Learning, Survival Analysis, Early Detection, Multi-modal Data
∗ Corresponding author: Mirza Faisal Beg, PhD, P. , Michael Smith Foundation for Health Research Scholar, School of Engineering
Science, Simon Fraser University, ASB 8857, 8888 University Drive, Phone: (778) 782-5696, Website: http://www2. Predicting Time-to-conversion for Dementia of Alzheimer’s
Type using Multi-modal Deep Survival Analysis
Ghazal Mirabnahrazam a,1, Da Ma a,b, Cédric Beaulac a,c, Sieun Lee a,d, Karteek Popuri a,e, Hyunwoo Lee f
, Jiguo
Cao g
, James E Galvin h
, Lei Wang i
, Mirza Faisal Beg a,*, and the Alzheimer’s Disease Neuroimaging Initiative 2
a School of Engineering, Simon Fraser University, Burnaby, BC, Canada
b School of Medicine, Wake Forest University, Winston-Salem, NC, USA
c Department of Mathematics and Statistics, University of Victoria, Victoria, BC, Canada
d Mental Health & Clinical Neurosciences, School of Medicine, University of Nottingham, Nottingham, United Kingdom
e Department of Computer Science, Memorial University of Newfoundland, St. John’s, NL, Canada
f Division of Neurology, Department of Medicine, University of British Columbia, Vancouver, BC, Canada
g Department of Statistics and Actuarial Science, Simon Fraser University, Burnaby, BC, Canada
h Comprehensive Center for Brain Health, Department of Neurology, University of Miami Miller School of Medicine, Miami, FL, USA
iPsychiatry and Behavioral Health, Ohio State University Wexner Medical Center, Columbus, OH, USA
Abstract
Dementia of Alzheimer’s Type (DAT) is a complex disorder influenced by numerous factors,
but it is unclear how each factor contributes to disease progression. The primary
goal of ADNI has been to test whether serial MRI, PET, other biological markers, and clinical and
neuropsychological assessment can be combined to measure the progression of mild cognitive
impairment (MCI) and early Alzheimer’s disease (AD). In addition, ADNI aims to provide researchers with
the opportunity to combine genetics with imaging and clinical data to help investigate mechanisms of the
disease. Keywords: Alzheimer’s Disease, Deep Learning, Survival Analysis, Early Detection, Multi-modal Data
∗ Corresponding author: Mirza Faisal Beg, PhD, P. , Michael Smith Foundation for Health Research Scholar, School of Engineering
Science, Simon Fraser University, ASB 8857, 8888 University Drive, Phone: (778) 782-5696, Website: http://www2. Keywords: Alzheimer’s Disease, Deep Learning, Survival Analysis, Early Detection, Multi-modal Data
∗ Corresponding author: Mirza Faisal Beg, PhD, P. , Michael Smith Foundation for Health Research Scholar, School of Engineering
Science, Simon Fraser University, ASB 8857, 8888 University Drive, Phone: (778) 782-5696, Website: http://www2. Keywords: Alzheimer’s Disease, Deep Learning, Survival Analysis, Early Detection, Multi-modal Data
∗ Corresponding author: Mirza Faisal Beg, PhD, P. , Michael Smith Foundation for Health Research Scholar, School of Engineering
Science, Simon Fraser University, ASB 8857, 8888 University Drive, Phone: (778) 782-5696, Website: http://www2. Data from various modalities were obtained and analyzed in the search for biomarkers that can
accurately diagnose DAT in its early stages. Magnetic resonance imaging (MRI), for example, is the most
widely used data modality for identifying specific structural changes in terms of atrophy in the brain
associated with DAT progression (Hua et al. Predicting Time-to-conversion for Dementia of Alzheimer’s
Type using Multi-modal Deep Survival Analysis
Ghazal Mirabnahrazam a,1, Da Ma a,b, Cédric Beaulac a,c, Sieun Lee a,d, Karteek Popuri a,e, Hyunwoo Lee f
, Jiguo
Cao g
, James E Galvin h
, Lei Wang i
, Mirza Faisal Beg a,*, and the Alzheimer’s Disease Neuroimaging Initiative 2
a School of Engineering, Simon Fraser University, Burnaby, BC, Canada
b School of Medicine, Wake Forest University, Winston-Salem, NC, USA
c Department of Mathematics and Statistics, University of Victoria, Victoria, BC, Canada
d Mental Health & Clinical Neurosciences, School of Medicine, University of Nottingham, Nottingham, United Kingdom
e Department of Computer Science, Memorial University of Newfoundland, St. John’s, NL, Canada
f Division of Neurology, Department of Medicine, University of British Columbia, Vancouver, BC, Canada
g Department of Statistics and Actuarial Science, Simon Fraser University, Burnaby, BC, Canada
h Comprehensive Center for Brain Health, Department of Neurology, University of Miami Miller School of Medicine, Miami, FL, USA
iPsychiatry and Behavioral Health, Ohio State University Wexner Medical Center, Columbus, OH, USA
Abstract
Dementia of Alzheimer’s Type (DAT) is a complex disorder influenced by numerous factors,
but it is unclear how each factor contributes to disease progression. Keywords: Alzheimer’s Disease, Deep Learning, Survival Analysis, Early Detection, Multi-modal Data
∗ Corresponding author: Mirza Faisal Beg, PhD, P. , Michael Smith Foundation for Health Research Scholar, School of Engineering
Science, Simon Fraser University, ASB 8857, 8888 University Drive, Phone: (778) 782-5696, Website: http://www2. Keywords: Alzheimer’s Disease, Deep Learning, Survival Analysis, Early Detection, Multi-modal Data
∗ Corresponding author: Mirza Faisal Beg, PhD, P. , Michael Smith Foundation for Health Research Scholar, School of Engineering
Science, Simon Fraser University, ASB 8857, 8888 University Drive, Phone: (778) 782-5696, Website: http://www2. Predicting Time-to-conversion for Dementia of Alzheimer’s
Type using Multi-modal Deep Survival Analysis
Ghazal Mirabnahrazam a,1, Da Ma a,b, Cédric Beaulac a,c, Sieun Lee a,d, Karteek Popuri a,e, Hyunwoo Lee f
, Jiguo
Cao g
, James E Galvin h
, Lei Wang i
, Mirza Faisal Beg a,*, and the Alzheimer’s Disease Neuroimaging Initiative 2
a School of Engineering, Simon Fraser University, Burnaby, BC, Canada
b School of Medicine, Wake Forest University, Winston-Salem, NC, USA
c Department of Mathematics and Statistics, University of Victoria, Victoria, BC, Canada
d Mental Health & Clinical Neurosciences, School of Medicine, University of Nottingham, Nottingham, United Kingdom
e Department of Computer Science, Memorial University of Newfoundland, St. John’s, NL, Canada
f Division of Neurology, Department of Medicine, University of British Columbia, Vancouver, BC, Canada
g Department of Statistics and Actuarial Science, Simon Fraser University, Burnaby, BC, Canada
h Comprehensive Center for Brain Health, Department of Neurology, University of Miami Miller School of Medicine, Miami, FL, USA
iPsychiatry and Behavioral Health, Ohio State University Wexner Medical Center, Columbus, OH, USA
Abstract
Dementia of Alzheimer’s Type (DAT) is a complex disorder influenced by numerous factors,
but it is unclear how each factor contributes to disease progression. 1 Introduction
Alzheimer's Disease (AD) or Dementia of Alzheimer’s Type (DAT) is a progressive
neurodegenerative condition characterized by psychiatric, cognitive and structural deteriorations that
accounts for 60% to 80% of all dementia cases. One out of every three seniors dies with Alzheimer’s
disease or another type of dementia, accounting for more deaths than breast and prostate cancer
combined (Alzheimer’s Association 2021). Combining MRI and
genetic features improved survival prediction over using either modality alone, but adding CDC to any
combination of features only worked as well as using only CDC features. Consequently, our study
demonstrated that using the current clinical procedure, which includes gathering cognitive test results,
can outperform survival analysis results produced using costly genetic or CSF data. 2008; Vemuri and Jack 2010; Popuri et al. Cerebrospinal Fluid (CSF) biomarkers such as abnormal amyloid and tau levels can reflect the intensity
of disease progression and have been extensively studied in relation to Alzheimer's disease (Olsson et
al. Therefore, here we use the time-dependent concordance index (Ctd-index) (Antolini, Boracchi,
and Biganzoli 2005), which estimates the probability that observations 𝑖 and 𝑗 are concordant
provided that they are comparable:
𝐶
𝑡𝑑 = 𝑃{ 𝑆̂(𝑇𝑖
| 𝑥𝑖
) < 𝑆̂(𝑇𝑗
| 𝑥𝑗) | 𝑇𝑖 < 𝑇𝑗
, 𝐷𝑖 = 1}. 2 Brier score
The Brier score (BS) (Brier 1950) is used to evaluate the accuracy of a predicted survival
function; it represents the average squared distances between the observed survival status, 𝑦𝑖 ∈
{0,1}, and the predicted survival probability, 𝑝̂𝑖 and is always a number between 0 and 1, with 0
being the best possible value. It is common in time-to-event data that
some individuals are not followed up to their event time for a variety of reasons, including leaving
the study or experiencing another event that prevents them from experiencing the event of
interest, resulting in censored times rather than event times. While many analyses ignore these
observations, one of the most important contributions of survival analysis methods is to account
for them. 2018) which extends the classic Cox proportional hazard model to predict,
analyze, and compare the time-to-conversion to DAT. Cox regression model (Cox 1972)
provides a semi-parametric specification of the hazard rate:
ℎ(𝑡|𝑥) = ℎ0
(𝑡)𝑒𝑥𝑝[𝑔(𝑥)], 𝑔(𝑥) = 𝛽
𝑇𝑥, (1)
where ℎ0
(𝑡) is a non-parametric baseline hazard, and exp[𝑔(𝑥)] is the hazard ratio or risk score. Therefore, here we use the time-dependent concordance index (Ctd-index) (Antolini, Boracchi,
and Biganzoli 2005), which estimates the probability that observations 𝑖 and 𝑗 are concordant
provided that they are comparable:
𝐶
𝑡𝑑 = 𝑃{ 𝑆̂(𝑇𝑖
| 𝑥𝑖
) < 𝑆̂(𝑇𝑗
| 𝑥𝑗) | 𝑇𝑖 < 𝑇𝑗
, 𝐷𝑖 = 1}. 2 Brier score
The Brier score (BS) (Brier 1950) is used to evaluate the accuracy of a predicted survival
function; it represents the average squared distances between the observed survival status, 𝑦𝑖 ∈
{0,1}, and the predicted survival probability, 𝑝̂𝑖 and is always a number between 0 and 1, with 0
being the best possible value. Section B in Supplementary Material
includes a comparison between this method and two other methods for handling missing values. Feature scaling has been performed on the data as a pre-processing step to ensure a consistent
set of feature ranges (Ma et al. To construct a non-linear version of the Cox model, the linear predictor 𝑔(𝑥) = 𝛽
𝑇𝑥 in
the relative risk function above is replaced by a 𝑔(𝑥) parametrized by a neural network. The
predictions are obtained by estimating the survival function, 𝑆̂(𝑡|𝑥) = exp[−𝐻̂(𝑡|𝑥)], where 𝐻̂(𝑡|𝑥)
is the cumulative hazard function, which is commonly used for specifying different survival
models, and is defined as: Survival analysis has received substantial recent attention in the machine learning
literature. The application of neural networks to survival analysis was first introduced by (Faraggi
and Simon 1995), where they extended the classical Cox proportional hazard model (Cox 1972)
by using a one hidden layer multilayer perceptron (MLP) to learn the relationship of the covariates
to the hazard function. We used 401 subjects with 63 features from MRI, genetic, and CDC (Cognitive tests,
Demographic, and CSF) data modalities in the Alzheimer’s Disease Neuroimaging Initiative (ADNI)
database. We used a deep learning-based survival analysis model that extends the classic Cox
regression model to predict time-to-conversion to DAT. Feature scaling has been performed on the data as a pre-processing step to ensure a consistent
set of feature ranges (Ma et al. Standardization (𝑥́ =
𝑥−𝑥̅
𝜎
) was performed on features with
either categorical or continuous data types, while binary features remained unchanged.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Deep learning based survival analysis models can predict time-to-conversion to Dementia of Alzheimer’s Type (DAT) by extending the classic Cox proportional hazard model to learn relationships between covariates and the hazard function. This allows AI systems to estimate when a subject will convert, accounting for censored individuals which traditional classification methods cannot do.
-
The system can perform multi-modal survival comparisons using metrics like time-dependent concordance index (Ctd-index) and Integrated Brier score (IBS) to determine the most predictive data modalities. This enables AI to select the optimal combination of MRI, genetic, and CDC features for prediction based on demonstrated performance superiority.
-
Feature importance analysis using Permutation Importance can identify which specific features contribute most positively or negatively to the model's performance across different feature sets. This allows AI to prune less useful features from a chosen set, leading to more parsimonious and robust predictive models.
-
The system can distinguish between non-progressive and progressive subjects by estimating survival probability over time using the GEN+MRI+CDC feature set. This capability allows for the prediction of disease progression trajectories, which is crucial for clinical risk stratification.
-
The AI can estimate predicted event times for progressive subjects using the model to determine if a subject is a DAT non-converter or converter based on their survival probability reaching 50 percent over a 10-year period. This offers clinical insight into the likelihood of future disease onset prior to actual diagnosis.
-
The system can be used to guide sequential data acquisition by determining the optimal order of data collection for at-risk patients. Specifically, it suggests that cognitive tests should be gathered first to determine health status, followed by genetic data and then subsequent follow-up visits for MRI and CSF data.
Abstract
Dementia of Alzheimer's Type (DAT) is a complex disorder influenced by numerous factors, but it is unclear how each factor contributes to disease progression. An in-depth examination of these factors may yield an accurate estimate of time-to-conversion to DAT for patients at various disease stages. We used 401 subjects with 63 features from MRI, genetic, and CDC (Cognitive tests, Demographic, and CSF) data modalities in the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. We used a deep learning-based survival analysis model that extends the classic Cox regression model to predict time-to-conversion to DAT. Our findings showed that genetic features contributed the least to survival analysis, while CDC features contributed the most. Combining MRI and genetic features improved survival prediction over using either modality alone, but adding CDC to any combination of features only worked as well as using only CDC features. Consequently, our study demonstrated that using the current clinical procedure, which includes gathering cognitive test results, can outperform survival analysis results produced using costly genetic or CSF data.
Related papers
- A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
- FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis