CoChat
Sign Up
Cited by 84Dementia and Cognitive Impairment ResearchAlzheimer's disease research and treatmentsArtificial Intelligence in Healthcare

Early Prediction Of Alzheimer’s Disease Dementia Based On Baseline Hippocampal MRI and 1-Year Follow-Up Cognitive Measures Using Deep Recurrent Neural Networks

Open access
Li, Hongming, Fan, Yong·2019·2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019)
DOI 10.1109/isbi.2019.8759397PMC PMC6892161
Open this paper in CoChat — free account

Abstract

Multi-modal biological, imaging, and neuropsychological markers have demonstrated promising performance for distinguishing Alzheimer's disease (AD) patients from cognitively normal elders. However, it remains difficult to early predict when and which mild cognitive impairment (MCI) individuals will convert to AD dementia. Informed by pattern classification studies which have demonstrated that pattern classifiers built on longitudinal data could achieve better classification performance than those built on cross-sectional data, we develop a deep learning model based on recurrent neural networks (RNNs) to learn informative representation and temporal dynamics of longitudinal cognitive measures of individual subjects and combine them with baseline hippocampal MRI for building a prognostic model of AD dementia progression. Experimental results on a large cohort of MCI subjects have demonstrated that the deep learning model could learn informative measures from longitudinal data for characterizing the progression of MCI subjects to AD dementia, and the prognostic model could early predict AD progression with high accuracy.

Integrity signals

Registry

Crossref · 10.1109/isbi.2019.8759397

Cites

21 of 21 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 21 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    All these models outperformed the baseline cognitive model (C-index 0.848).

    “All these prediction models had better prediction performance than the model built upon the baseline cognitive measures (a C-index value of 0.848), while the LSTM based prediction models were significantly better (p<0.05).”
    Quote in the paperSubstantial claim
  • Findingown data

    The 12-month combined longitudinal cognitive and baseline imaging model achieved C-index 0.901.

    “When both longitudinal cognitive measures and baseline imaging measures were used, prediction models achieved C-index values of 0.901 and 0.889 at 12 and 6 months respectively, significantly better (p<0.006) than a model built on baseline cognitive measures and imaging measures (a C-index value of 0.866).”
    Quote in the paperSubstantial claim
  • Findingown data

    For ADNI-GO&2 subjects, the three-time-point model trained on ADNI-1 cognitive trajectories reached C-index 0.896.

    “When the prediction models built upon ADNI-1 longitudinal cognitive data of 3 and 2 time points applied to ADNI-GO&2 subjects, they achieved C-index values of 0.896 and 0.873 respectively, better than the prediction models built upon single-visit data at 12 and 6 months (p=0.019 and 0.552 respectively).”
    Quote in the paperSubstantial claim
  • Findingown data

    The 6-month combined longitudinal cognitive and baseline imaging model achieved C-index 0.889.

    “When both longitudinal cognitive measures and baseline imaging measures were used, prediction models achieved C-index values of 0.901 and 0.889 at 12 and 6 months respectively, significantly better (p<0.006) than a model built on baseline cognitive measures and imaging measures (a C-index value of 0.866).”
    Quote in the paperSubstantial claim
  • Findingown data

    The three-time-point model outperformed the single-visit 12-month model (p=0.019).

    “When the prediction models built upon ADNI-1 longitudinal cognitive data of 3 and 2 time points applied to ADNI-GO&2 subjects, they achieved C-index values of 0.896 and 0.873 respectively, better than the prediction models built upon single-visit data at 12 and 6 months (p=0.019 and 0.552 respectively).”
    Quote in the paperSubstantial claim
  • Classified as probably not a substantial claim. 3
  • May combine two claims. 5

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductnone given

    For pMCI subjects who converted at a visit or sMCI subjects lacking data at a visit, only earlier-visit data were used.

    “It is worth noting that if a pMCI subject converted to AD at Visit-ID or data at Visit-ID were not available for a sMCI subject, only data at visits before the Visit-ID were used.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisenone given

    In classification-based AD dementia prediction, performance depends on the follow-up cut-off defining progressive and stable MCI, and the resulting groups are typically heterogeneous regardless of threshold.

    “However, the classification performance is dependent on a cut-off threshold of follow-up duration that is used to define pMCI and sMCI. Moreover, the cohorts of pMCI and sMCI subjects are typically heterogeneous regardless of the threshold used.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    The learned representations can encode temporal dynamics of longitudinal cognitive measures and describe progression trajectories without explicit assumptions about the underlying longitudinal process.

    “These representations could encode temporal dynamics of longitudinal cognitive measures and characterize the progression trajectory of MCI subjects without any explicit assumption regarding the longitudinal process behind the measures.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Longitudinal prediction models generally demand synchronized measurement times across subjects.
    • Rests on an assumption: Longitudinal studies commonly suffer from missing data.

  • May combine two claims.

    Study design and conductown data

    The Cox regression model was trained on ADNI-1 MCI subjects and evaluated on ADNI-GO&2 MCI subjects.

    “Cox regression model was built on data of MCI subjects of ADNI-1, and its prognostic performance was evaluated based on the data of MCI subjects of ADNI-GO&2.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisecitation

    MFPC scores have been used for longitudinal marker representation and missing/irregular data, but their latent-process assumption may be inappropriate for different marker types.

    “Multivariate functional principal component (MFPC) scores [12] have been used to represent longitudinal makers for handling missing or irregular data. However, certain assumption is adopted in MFPC to model the latent longitudinal process, which may not be appropriate for different types of markers.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [12] Li K, O’Brien R, Lutz M, Luo S, and Alzheimer’s Disease Neuroimaging I, “ A prognostic model of Alzheimer’s disease relying on multiple longitudinal measures and time-to-event data,” Alzheimers Dement, vol. 14, no. 5, pp. 644– 651, 5 2018. 29306668 PMC5938096 source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    LSTM-based RNNs have made notable advances in sequence modeling, including machine translation and functional MRI modeling.

    “Recently, deep learning techniques built upon recurrent neural networks (RNNs) with a long short term memory (LSTM) [17] structure have achieved remarkable advances in sequence modeling, such as machine translation and functional MRI modeling [18, 19], indicating RNNs might be better tools for characterizing longitudinal data.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [17] Hochreiter S and Schmidhuber J, “ Long short-term memory,” Neural computation, vol. 9, no. 8, pp. 1735– 1780, 1997. 9377276 source

      Not checked in the free run.

    • cites [18] Li H and Fan Y, “ Brain Decoding from Functional MRI Using Long Short-Term Memory Recurrent Neural Networks,” in Medical Image Computing and Computer Assisted Intervention – MICCAI 2018, Cham, 2018, pp. 320– 328: Springer International Publishing. PMC6180332 30320311 source

      Not checked in the free run.

    • cites [19] Li H and Fan Y, “ Identification of Temporal Transition of Functional States Using Recurrent Neural Networks from Functional MRI,” in Medical Image Computing and Computer Assisted Intervention – MICCAI 2018, Cham, 2018, pp. 232– 239: Springer International Publishing. PMC6180329 30320310 source

      Not checked in the free run.

6 of 6

  • The paper relies on this without support.

    It is still hard to forecast at baseline which MCI patients will develop AD dementia and when.

    “However, it remains difficult to early predict when and which mild cognitive impairment (MCI) individuals will convert to AD dementia.”
    “it remains difficult at baseline to predict which and when MCI individuals will convert to AD dementia.”

  • The paper relies on this without support.

    Longitudinal prediction models generally demand synchronized measurement times across subjects.

    “Most longitudinal data based prediction models require different subjects to have data at the same time points.”

  • The paper relies on this without support.

    Longitudinal studies commonly suffer from missing data.

    “However, missing data is a ubiquitous problem in longitudinal studies.”
3 of 3

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Prediction of Mild Cognitive Impairment or Alzheimer's disease based on ADNI dataset is frequently investigated, but predictive accuracy varies, and many studies do not discuss which input modalities or features contributed most to prediction.

    “For example, the prediction of Mild Cognitive Impairment or Alzheimer’s disease based on the ADNI dataset is frequently investigated27–32. But their predictive accuracy varies, and many studies do not discuss which input modalities or features contributed most to the prediction.”

    From Orchestrating explainable artificial intelligence for multimodal and longitudinal data in medical imaging (2024) · cites this paper as [30]

    Split into 2 atoms — 1 confirmed, 0 partially supported, 1 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

    Checked by a reasoning model — may be wrong.

    Experimental — accuracy not yet measured

  • Four out of the six studies that applied neural networks for feature extraction simply concatenated the extracted imaging features with clinical features for their fusion strategy.

    “Four out of the six studies that applied neural networks for feature extraction simply concatenated the extracted imaging features with clinical features for their fusion strategy26–29.”

    From Fusion of medical imaging and electronic health records using deep learning: a systematic review and implementation guidelines (2020) · cites this paper as [28]

    The Methods and Materials section (“Prognostic modeling”; page unavailable) says that imaging features were extracted using a deep-learning framework and combined with LSTM-encoded cognitive measures in a Cox regression model. This supports the paper’s inclusion as an example of imaging–clinical feature fusion, but our copy does not specify that the features were simply concatenated. The four-of-six count is the citing paper’s tally across several studies, not a finding this individual source establishes. (LLM, conf=0.78)

    Experimental — accuracy not yet measured

  • Fusion strategies were often applied to the diagnosis and prediction of Alzheimer's disease.

    “For example, fusion strategies were often applied to the diagnosis and prediction of Alzheimer’s disease25,28,33,36,41.”

    From Fusion of medical imaging and electronic health records using deep learning: a systematic review and implementation guidelines (2020) · cites this paper as [28]

    The Methods and Materials (Prognostic modeling) and Results (Experimental results; page not provided) describe combining cognitive measures with hippocampal MRI features to predict progression to Alzheimer’s dementia. This supports the paper as an example of fusion for prediction, but it does not show its fusion strategy being used for diagnosis or establish how often such strategies were applied. (LLM, conf=0.82)

    Experimental — accuracy not yet measured

  • Six of the seven early fusion studies that compared fusion to single modality models showed an improvement in performance when using fusion, and the remaining one achieved the same performance but reduced standard deviation.

    “Six of these studies showed an improvement in performance when using fusion25,26,28,29,31,33, and the remaining one achieved the same performance but reduced standard deviation27, alluding to a model with better stability.”

    From Fusion of medical imaging and electronic health records using deep learning: a systematic review and implementation guidelines (2020) · cites this paper as [28]

  • Unless there is a compelling reason otherwise, outputs from linear layers of a CNN are usually effective feature representations of the original image.

    “It is worth noting, that unless there is a compelling reason for using such an approach, outputs from linear layers of a CNN are usually effective feature representations of the original image28,29,31.”

    From Fusion of medical imaging and electronic health records using deep learning: a systematic review and implementation guidelines (2020) · cites this paper as [28]

    The Methods section describes an LSTM autoencoder that learns representations of longitudinal cognitive measures, and the Results section discusses combining those representations with MRI-derived features. Searches for CNNs, linear layers, and image feature representations found no passage assessing whether outputs from a CNN’s linear layers are usually effective representations of an image. (LLM, conf=0.74)

    Experimental — accuracy not yet measured

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. li, 0, A deep learning prognostic model for early prediction of Alzheimer&#x2019;s disease based on hippocampal MRI data, presented at the Alzheimer&#x2019;s Association International Conference 2018

Cited by

Sorted by citations

Related

Authors
Li, H., & Fan, Y.
Affiliations
University of Pennsylvania (US)
ORCID
Li, Hongming
Venue
2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019)
Publisher
IEEE
Published
2019-04
Volume
2019
Pages
368-371
DOI
10.1109/isbi.2019.8759397
PMC
PMC6892161
OpenAlex
W2964078365
Cited by
83
References
21
Citations by year
2026: 7 · 2025: 17 · 2024: 11 · 2023: 19 · 2022: 12 · 2021: 10 · 2020: 6 · 2019: 1
FWCI
21.06
Type
conference paper
Language
en
Open access
Licence
Locations
  • IEEE International Symposium on Biomedical Imaging (ISBI) (publishedVersion)
  • PubMed (publishedVersion)
  • PubMed Central (acceptedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • National Institutes of Health (1s10od023495-01, AG054409)
  • National Institute on Aging (RF1AG054409)
  • National Cancer Institute (R21CA223358)
  • National Institute of Biomedical Imaging and Bioengineering (R01EB022573)
Topics
  • Health Sciences › Medicine › Psychiatry and Mental health › Dementia and Cognitive Impairment Research
  • Health Sciences › Medicine › Physiology › Alzheimer's disease research and treatments
  • Health Sciences › Health Professions › Health Information Management › Artificial Intelligence in Healthcare
Keywords
Alzheimer's disease, dementia prediction, mild cognitive impairment, MCI conversion, hippocampal MRI, deep recurrent neural networks, early prediction, disease progression, prognostic model

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-01.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-01.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review