Pith. sign in

REVIEW 3 major objections 6 minor 68 references

From population norms to personalized trajectories: interpretable Bayesian forecasting for cognitive decline

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Cognitive decline can be spotted years early by comparing a person's scores with their own predicted baseline, not population norms.

desk verdict A transparent, well-evaluated framework whose headline early-detection numbers need a grain of salt (HRS endpoint overlap, unvalidated age-adjustment assumption) but the ADNI validation makes it a genuine contribution. read the letter →

arxiv 2608.03877 v1 pith:EWXPXH5L submitted 2026-08-04 cs.ET

Maria Sahakyan , Brita Elvev{aa}g This is my paper
classification cs.ET
keywords cognitivedeclineBayesianupdatingExplainableBoostingMachinepersonalizedbaselineearlydetectionlongitudinalmonitoringHealthandRetirementStudyADNI
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that cognitive decline should be monitored against each person's own expected trajectory rather than a fixed population cutoff. PRISM first predicts a personalized baseline memory score from routine demographic and health data using an interpretable additive model, then updates that prediction visit by visit with Bayesian inference that weights recent scores more heavily. Decline is flagged only when both the estimated slope and the current level fall below an age-adjusted personal anchor. In two longitudinal cohorts, the framework flagged 31% and 41% of sustained decliners before their first categorical transition to worse cognitive status, with median lead times of 6 and 2 years, and by the time of worsening it had flagged 68% and 56%. If correct, routine cognitive testing could yield an interpretable, uncertainty-aware early-warning signal even when a patient's own history is still short.

What carries the argument

The load-bearing mechanism is a three-stage pipeline. Stage 1 uses an Explainable Boosting Machine, an interpretable generalized additive model, to predict each person's baseline memory score from routine baseline variables, producing a personalized prior. Stage 2 updates this prior with each new observed score via Normal-Normal conjugate Bayesian updating, with older observations exponentially down-weighted by e^{-λΔt}; the tuning parameter λ controls how quickly past scores lose influence. Stage 3 estimates a Bayesian slope with a zero-centered prior, giving a posterior probability of decline P(slope<0). A three-tier detection rule then flags decline when both the slope probability and a b

What would settle it

Take a group of cognitively stable older adults, test them with the same memory test at every visit for 10 or more years, and run PRISM with the same settings. If retest practice effects raise later scores, the age-adjusted anchor will drift below the true stable trajectory, and the false-alert rate among this stable group should exceed the 30% ceiling used in tuning, which would show the early-detection signal is partly an artifact of the age assumption.

Watch

Extended reading notes

Core claim

On its own terms, the discovery is that a population-informed but person-specific baseline, refined by Bayesian updating with exponential temporal decay, can detect emerging cognitive decline before a person crosses a categorical threshold of impairment. The paper reports that among participants who later sustained cognitive worsening, PRISM identified 31% in the Health and Retirement Study and 41% in the Alzheimer's Disease Neuroimaging Initiative before the first transition visit, with median lead times of six and two years; by the transition visit, 68% and 56% had been flagged. PRISM also produced lower one-step-ahead forecasting error than a demographic norm or a cumulative personal aver

Load-bearing premise

The system assumes that the age-related decline curve learned by comparing different people's scores at different ages matches the true path each individual's healthy brain would follow; if birth-cohort differences or retest practice effects bend that curve, the personal anchor drifts and the flag positions and lead times shift.

Editorial extensions

If this is right

  • A clinician could receive a personalized expected score at the first visit, so monitoring need not wait for a long personal history.
  • Each new test score updates the forecast and its uncertainty, making the system suitable for visit-by-visit clinical tracking.
  • Population norms could be replaced or complemented by individualized age-adjusted anchors in screening programs.
  • Earlier identification, years before categorical transition in a subset of patients, could widen the window for preventive interventions and trial enrollment.
  • The same three-stage structure can be ported to other cognitive domains or slowly changing health measures by retuning thresholds.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The lead-time numbers likely depend on how much the monitored memory test overlaps with the outcome definition; an independent endpoint built from different tests would be a stronger test of the early-detection claim.
  • Because the age adjustment is learned cross-sectionally, applying PRISM to cohorts with different birth-cohort cognitive trends or strong retest practice effects could shift both the anchor and the reported lead times; comparing against a decades-long within-person follow-up would quantify this.
  • The global decay parameter λ could become person- or condition-specific; the paper reports a cohort-level optimum, leaving open whether adaptive decay would catch abrupt declines faster without raising false alarms.
  • A prospective deployment where PRISM alerts trigger additional clinical assessment would test whether the earlier warnings change patient management or outcomes, something the retrospective evaluation cannot show.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper introduces PRISM, a three-stage interpretable framework for individualized longitudinal monitoring of cognitive decline. Stage 1 uses an Explainable Boosting Machine (EBM) trained on baseline demographic/health/functional variables to produce a personalized cognitive baseline; Stage 2 updates this baseline sequentially with observed cognitive scores via normal–normal Bayesian updating with exponential temporal decay; Stage 3 estimates a posterior probability of a negative slope, P(decline), using a zero-centered prior on the slope. A three-tier detection rule combines P(decline) with a below-anchor level check to flag decline relative to an age-adjusted personal anchor. The authors evaluate PRISM in the Health and Retirement Study (N=30,664; 70/15/15 split) and externally validate in ADNI (N=1,866; 5-fold CV), comparing one-step-ahead forecasting to demographic-norm and cumulative-average baselines, and per-visit discrimination to a linear mixed-effects model. The headline results are that, among sustained decliners, PRISM flagged 31% (HRS) and 41% (ADNI) before study-defined cognitive worsening, with 68% and 56% flagged by the transition visit, and median lead times of 6 and 2 years, respectively.

Significance. If the results hold, PRISM is a practically useful, transparent early-warning system for cognitive decline that can operate from sparse longitudinal histories using routinely collected variables. The paper has several concrete strengths: it releases code; the evaluation uses held-out test data in HRS and an independent external cohort (ADNI) with clinician-assigned diagnoses and a different memory test; the Bayesian updating and detection rules are described precisely with equations; and the comparison with an LME baseline addresses a clinically relevant question. The central novelty—combining a population-informed EBM prior with sequential Bayesian updating and personalized anchors—is sensible and well motivated. The main risks are the unvalidated cross-sectional-to-longitudinal age adjustment, the partial overlap between the monitored signal and the HRS endpoint, and the absence of reported support for the abstract's median lead-time claims. These are fixable with additional analyses and transparent reporting, but they are load-bearing for the paper's main claims.

major comments (3)
  1. [Supplementary Note 2.1, Eq. (S20)] The age-adjusted baseline and personal anchor are constructed by shifting the EBM baseline by [f_age(alpha_i + Delta t) - f_age(alpha_i)], where f_age is the EBM age shape function learned cross-sectionally from baseline data. This assumes the between-person age gradient equals each individual's true within-person healthy-aging trajectory. The claim that PRISM detects decline 'beyond that expected in healthy aging' and the threshold/lead-time statistics both depend on this assumption. If f_age overstates age-related decline, the anchor drifts downward too quickly and genuine decliners must drop further to be flagged; if it understates decline, normal aging may be misclassified. The paper provides no sensitivity analysis for this assumption, and the ADNI validation does not address it because the same EBM age-adjustment construction is reused. Please add a sensitivity analysis that replac
  2. [Abstract and Results (Figure 3B)] The abstract reports median lead times of 6 years (HRS) and 2 years (ADNI), but the Results section and Figure 3B only report the cumulative percentages of sustained decliners flagged at one visit before transition (31%/41%) and by the transition visit (68%/56%). No lead-time distribution, median, or interquartile range appears in the main text or figures. Since median lead time is a headline quantitative claim, it must be explicitly reported and defined (visits vs. calendar years), preferably as a Kaplan–Meier-style cumulative detection curve with a table of lead-time quantiles. Without this, the claimed lead times cannot be verified from the presented results.
  3. [Methods, HRS outcome definition and Discussion limitations] In HRS, the monitored longitudinal signal is the word-recall composite, which contributes 20 of the 27 points to the Langa–Weir total cognition score that defines the worsening endpoint. PRISM's early detection in HRS is therefore partly tautological: a decline in the monitored signal directly moves the endpoint. The authors acknowledge this and point to the ADNI validation, which uses clinician-assigned diagnoses and a different memory test (RAVLT delayed recall). This mitigation is real but incomplete, because RAVLT performance is part of the neuropsychological information used in ADNI diagnostic classification. I recommend a sensitivity analysis that defines the HRS endpoint using only the non-word-recall components of the 27-point score (serial subtraction and backward counting), or at least reports the detection statistics restricted to transitions not driven solely by word-recall c
minor comments (6)
  1. [Methods, HRS eligibility] Duplicate phrase: 'were not were not classified as having dementia at baseline' should read 'were not classified as having dementia at baseline.' Also, 'altering word lists' should be 'alternating word lists.'
  2. [Figure 2 caption] Panel B shows 'HHIDPM' while the other panels use 'HHIDPN'; please confirm this is not a typographical inconsistency.
  3. [Methods, ADNI split] 'participants were derived into training, validation and test sets' should be 'divided into.'
  4. [Discussion] The sentence 'beyond that exceed what would be expected in healthy aging' is grammatically awkward; suggest 'beyond what would be expected in healthy aging.'
  5. [Supplementary Note 2.7, Eq. (S34)] The composite decline score is used for the AUC analyses but is not defined in the main Methods; please introduce it at first use in the main text.
  6. [Figure 1C] The MAE plot would benefit from explicit confidence intervals or a statistical comparison of the three methods at each visit, rather than only point estimates, to support the claim that PRISM's lower MAE is not within sampling variability.

Circularity Check

1 steps flagged · score 4.0 of 10

HRS early-detection result is partly self-definitional because the monitored word-recall signal is a component of the Langa-Weir endpoint; ADNI provides an independent validation so the central claim is not fully circular.

  1. self definitional [Discussion, Limitations; Methods, Outcome definition (HRS)]
    "First, the HRS reference outcome partly overlaps, creating overlaps with the word recall measure monitored by PRISM, although this concern is mitigated by the ADNI validation using clinician-assigned diagnostic worsening based on broader clinical information and a different memory measure."

    The HRS endpoint is the Langa-Weir classification, whose 27-point total = word recall (0-20) + serial subtraction (0-5) + backward counting (0-2). PRISM monitors word recall and flags decline when word recall falls below an anchor with negative slope. A decline in the monitored word recall directly lowers the composite defining worsening, so the HRS 'early detection before worsening' statistic is partially forced by construction: the signal is an additive component of the outcome. The paper explicitly acknowledges this. ADNI uses a different memory measure and clinician diagnosis, so the central claim retains independent content; however, HRS headline numbers (31%, 68%) are partly self-definitional.

full rationale

The only substantive circularity is the partial overlap between the monitored signal and the HRS endpoint, which the paper itself discloses and mitigates with ADNI. This is a genuine definitional coupling: a negative word-recall trajectory both triggers PRISM and lowers the Langa-Weir total used to define sustained decline. No other load-bearing step reduces to its inputs. The temporal decay parameter and detection thresholds are tuned on validation splits and applied to held-out test data, so the test statistics are not fitted-by-construction. The EBM age adjustment (Eq. S20) is an identifying assumption equating cross-sectional age gradients with within-person healthy aging; it is a correctness risk, not a circularity, because f_age is learned from baseline data and is not defined in terms of the decline endpoint. Reference [8] is a self-citation (co-author Elvevåg) used for conceptual framing of forecasting versus classification, but it is not load-bearing for any mathematical claim. The ADNI external validation uses a different memory measure and clinical diagnosis, giving the central early-detection claim independent support. Overall score 4: one partially circular evaluation component, central claim still independently supported.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

PRISM introduces no new ontological entities; its load is carried by fitted parameters (lambda, tier thresholds, margins), a set of distributional and endpoint-validity assumptions, and one strong modeling premise: that the cross-sectionally learned EBM age function equals the within-person aging trajectory. The HRS endpoint overlap between monitored signal and reference outcome is the main structural circularity, partially mitigated by ADNI.

free parameters (5)
  • Temporal decay rate lambda = 0.20 (HRS); 0.96 +/- 0.08 (ADNI)
    Selected on validation by minimizing one-step-ahead MAE (Eq. S29); controls how quickly older observations are down-weighted in the Bayesian level update (Eq. S8).
  • Tier 1 threshold pimm, Tier 2 threshold psus, confirmation length nconfirm = Age-band-specific values selected on validation
    Tuned per age band with a hierarchical objective maximizing a saturating early-detection lead-time reward under a 0.70 hard-negative specificity floor (Supp Note 2.4.2). The headline detection statistics inherit these choices.
  • Margin m and minimum gap Delta_min per age band = m=0.5 (50-74), m=0.0 (75+); Delta_min=0.20 (50-74), 0.10 (75+)
    Pre-specified by hand (Supp Table 2) and subtracted from the age-adjusted anchor in the below-anchor check (Eq. S23); chosen, not fitted, but they shift flag sensitivity.
  • Test-retest reliability r = 0.70 (fixed, both cohorts)
    Fixed constant entering SEM = SD*sqrt(1-r) (Eq. S7); supported by literature (0.76-0.80 in ADNI) and an internal HRS correlation of 0.74, but a chosen value that sets the observation precision.
  • Optimization hyperparameters = 0.70 specificity floor; saturation rate k and tolerances not reported
    Hyperparameters of the detection-rule optimization that directly shape the sensitivity/specificity trade-off behind the early-detection headline; k and the epsilon tolerances are not stated in the paper.
assumptions (6)
  • standard math Normal-normal conjugacy for level and slope inference
    Stage 2 and Supp Note 1.2 (Eqs. S5-S11); slope posterior Eqs. S14-S18 assume Normal prior and Normal observation noise.
  • domain assumption EBM residual standard deviation sigma0 as prior variance
    Eq. S4 sets prior variance to in-sample residual SD of the EBM on training data; treats model misspecification as symmetric zero-mean noise.
  • domain assumption Zero-centered slope prior with empirical sigma_beta
    Eq. S14 centers the slope prior on no change; stabilizes early slopes but makes P(decline) near 0.5 for short histories by construction.
  • domain assumption Cross-sectional EBM age function represents within-person healthy aging
    Eq. S20 adjusts the baseline and personal anchor using the cross-sectionally learned EBM age shape function; cohort and practice effects would bias the anchor.
  • domain assumption Sustained-decline endpoint and classifications are valid reference truth
    Sustained-decline endpoint (Eq. S28) and the Langa-Weir (HRS) and clinical diagnosis (ADNI) classifications are treated as reference truth; in HRS the monitored word-recall score is 20 of the 27 points of that classification.
  • domain assumption Constant test-retest reliability across time and individuals
    SEM is assumed stationary within and across persons; no heteroscedasticity or practice-effect trend is modeled.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From population norms to personalized trajectories: interpretable Bayesian forecasting for cognitive decline." pith.science (2026). https://pith.science/paper/EWXPXH5L

@misc{pith2026260803877,
  author       = {Pith},
  title        = {Pith review of: From population norms to personalized trajectories: interpretable Bayesian forecasting for cognitive decline},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EWXPXH5L}},
  note         = {Machine review of arXiv:2608.03877}
}
read the original abstract

The timing and rate of cognitive decline vary substantially between individuals, limiting the ability of fixed population-level thresholds to determine whether a new observation represents meaningful change for an individual. We developed Personalized Risk Inference via Sequential Monitoring (PRISM), an interpretable framework for individualized longitudinal forecasting of cognitive decline. PRISM estimates a personalized cognitive baseline from routinely collected demographic, health, and functional variables using an Explainable Boosting Machine, then updates this expectation through Bayesian inference with temporal decay as cognitive scores accrue. Decline is evaluated relative to an age-adjusted personal anchor, with uncertainty quantified through posterior probabilities. We evaluated PRISM in 30,664 adults from the Health and Retirement Study and externally validated it in 1,866 Alzheimer's Disease Neuroimaging Initiative participants. Forecasting performance was compared with demographic-norm and cumulative-average baselines, and discrimination with a linear mixed-effects model. PRISM identified emerging decline before study-defined cognitive worsening in 31% of sustained decliners in the Health and Retirement Study and 41% in the Alzheimer's Disease Neuroimaging Initiative, with median lead times of 6 and 2 years, respectively. By the time of worsening, 68% and 56% had been identified. PRISM also achieved lower forecasting error than demographic-norm and cumulative-average baselines and distinguished worsening from stable trajectories better than a linear mixed-effects model, particularly early in follow-up. PRISM enables earlier, interpretable, uncertainty-aware detection of cognitive decline relative to each individual's expected trajectory using routinely collected data. It may support closer monitoring and timely assessment when personal longitudinal history is limited.

Figures

Figures reproduced from arXiv: 2608.03877 by the authors.

Figure 1
Figure 1. Longitudinal behavior and forecasting performance of the PRISM framework in the Health and Retirement Study (HRS) test set (N = 4,600; 1996–2020). (A) Bayesian adaptation from population prior to individual history. Relative contribution of the EBM-derived population prior and individualized longitudinal history to the posterior estimate across visits. Individual history rapidly becomes dominant after the first obse… view at source ↗
Figure 2
Figure 2. Representative cognitive trajectories from the HRS test set illustrating the three PRISM decline-detection tiers. Each panel shows one individual’s longitudinal word-recall trajectory together with the corresponding PRISM one-step-ahead Bayesian forecast. Observed visit scores are marked according to the Langa–Weir cognitive classification at that visit: Normal cognition, CIND/MCI, or dementia. The Bayesian forecast… view at source ↗
Figure 3
Figure 3. Transition-centered alignment of longitudinal trajectories and cumulative early detection by PRISM. (A) Transition-centered alignment procedure used for longitudinal analyses. Individual trajectories were aligned relative to the first visit at which worsened cognitive status was observed, defined as the transition visit (visit 0). Negative visit numbers represent assessments preceding the transition, whereas positiv… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Overview of the longitudinal cognitive decline detection framework across the HRS and ADNI cohorts. (A) Cohort construction, longitudinal data structure, and personalized baseline estimation. Participants with baseline dementia, age < 50 years, proxy cognitive assessme…
Figure 5
Figure 5. Figure 5: Illustrative example of the three-stage framework for personalized cognitive decline detection, shown for three individuals from the Health and Retirement Study (HRS). (A) Per￾sonalized baseline estimation via EBM. An Explainable Boosting Machine (EBM) learns feature-s…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 62 canonical work pages

  1. [1]

    Alzheimer’s disease: epidemiology and clinical progression.Neurology and therapy, 11(2): 553–569, 2022

    Amir Abbas Tahami Monfared, Michael J Byrnes, Leigh Ann White, and Quanwu Zhang. Alzheimer’s disease: epidemiology and clinical progression.Neurology and therapy, 11(2): 553–569, 2022

    work page 2022

  2. [2]

    The challenging concept of preclinical alzheimer’s disease.Revue Neurologique, 181(9):881–892, 2025

    D Wallon and A Garnier-Crussard. The challenging concept of preclinical alzheimer’s disease.Revue Neurologique, 181(9):881–892, 2025

    work page 2025

  3. [3]

    Alzheimer’s disease–why we need early diagnosis

    Jill Rasmussen and Haya Langerman. Alzheimer’s disease–why we need early diagnosis. Degenerative neurological and neuromuscular disease, pages 123–130, 2019

    work page 2019

  4. [4]

    Reisa A Sperling, Paul S Aisen, Laurel A Beckett, David A Bennett, Suzanne Craft, Anne M Fagan, Takeshi Iwatsubo, Clifford R Jack Jr, Jeffrey Kaye, Thomas J Montine, et al. Toward defining the preclinical stages of alzheimer’s disease: Recommendations from the national institute on aging-alzheimer’s association workgroups on diagnostic guidelines for alzh...

    work page 2011

  5. [5]

    The best treatment is prevention: preven- tion of cognitive decline and dementia–current state, gaps and next steps–.Neurological Research and Practice, 8(1):30, 2026

    Kathrin Reetz, Inga Liepelt-Scarfone, Alexa H¨ ager, Agnes Fl¨ oel, J¨ org B Schulz, and Demen- tia Commission of the German Society for Neurology (DGN) von Arnim Christine Dodel Richard Duning Thomas Kaluza Luisa M¨ unte Thomas Kleineberg Nina Saur Dorothee Nellessen Nils Schwichtenberg Svenja Christine. The best treatment is prevention: preven- tion of ...

    work page 2026

  6. [6]

    Gulsah Hancerliogullari Koksalmis, Bulent Soykan, Laura J Brattain, and Hsin-Hsiung Huang. Artificial intelligence for personalized prediction of alzheimer’s disease progression: A survey of methods, data challenges, and future directions.arXiv preprint arXiv:2504.21189, 2025

    work page Pith review arXiv 2025

  7. [7]

    The cognitive reserve hypothesis: a longitudinal examination of age-associated declines in reasoning and processing speed.Developmental psychology, 45(2):431, 2009

    Elliot M Tucker-Drob, Kathy E Johnson, and Richard N Jones. The cognitive reserve hypothesis: a longitudinal examination of age-associated declines in reasoning and processing speed.Developmental psychology, 45(2):431, 2009

    work page 2009

  8. [8]

    The challenges of implementing hybrid baselines for the interpretation of longitudinal behavioral data from individuals.npj Digital Medicine, 9 (1):331, 2026

    Sandra Anna Just, Enrico Tedeschi, Einar Holsbø, Karl Øyvind Mikalsen, Lars Ailo Bongo, Philipp Homan, and Brita Elvev ˚ ag. The challenges of implementing hybrid baselines for the interpretation of longitudinal behavioral data from individuals.npj Digital Medicine, 9 (1):331, 2026

    work page 2026

Show all 68 references
  1. [9]

    Manuel C Voelkle, Annette Brose, Florian Schmiedek, and Ulman Lindenberger. Toward a unified framework for the study of between-person and within-person structures: Building a bridge between two research paradigms.Multivariate Behavioral Research, 49(3):193–213, 2014

    2014

  2. [10]

    Lack of group-to-individual generalizability is a threat to human subjects research.Proceedings of the National Academy of Sciences, 115(27):E6106–E6115, 2018

    Aaron J Fisher, John D Medaglia, and Bertus F Jeronimus. Lack of group-to-individual generalizability is a threat to human subjects research.Proceedings of the National Academy of Sciences, 115(27):E6106–E6115, 2018

    2018

  3. [11]

    Expanding the use of brief cognitive assessments to detect suspected early-stage cognitive impairment in primary care.Alzheimer’s & Dementia, 19(9):4252–4259, 2023

    Soeren Mattke, Donnie Batie, Joshua Chodosh, Kristen Felten, Ellen Flaherty, Nicole R Fowler, Fred A Kobylarz, Kelly O’Brien, Russ Paulsen, Anne Pohnert, et al. Expanding the use of brief cognitive assessments to detect suspected early-stage cognitive impairment in primary car...

    2023

  4. [12]

    Accessible analysis of longitudinal data with linear mixed effects models.Disease Models & Mechanisms, 15(5): dmm048025, 2022

    Jessica I Murphy, Nicholas E Weaver, and Audrey E Hendricks. Accessible analysis of longitudinal data with linear mixed effects models.Disease Models & Mechanisms, 15(5): dmm048025, 2022

    2022

  5. [13]

    Shreya K Rajagopal, Adriene M Beltz, Benjamin M Hampstead, and Thad A Polk. Estimating individual trajectories of structural and cognitive decline in mild cognitive impairment for early prediction of progression to dementia of the alzheimer’s type.Scientific Reports, 14(1):12906, 2024

    2024

  6. [14]

    Evaluating the performance of bayesian and frequentist approaches for longitudinal modeling: application to alzheimer’s disease.Scientific Reports, 12(1):14448, 2022

    Agnes Perez-Millan, Jos´ e Contador, Ra´ ul Tudela, Aida Ninerola-Baizan, Xavier Setoain, Albert Llado, Raquel Sanchez-Valle, and Roser Sala-Llonch. Evaluating the performance of bayesian and frequentist approaches for longitudinal modeling: application to alzheimer’s disease....

    2022

  7. [15]

    That blup is a good thing: the estimation of random effects.Statistical science, pages 15–32, 1991

    George K Robinson. That blup is a good thing: the estimation of random effects.Statistical science, pages 15–32, 1991

    1991

  8. [16]

    Yi Li, Yaning Yang, Xu Steven Xu, Min Yuan, Alzheimer’s Disease Neuroimaging Initiative, et al. Bias correction for multiple covariate analysis using empirical bayesian estimation in mixed-effects models for longitudinal data.Computational Biology and Chemistry, 99: 107697, 2022

    2022

  9. [17]

    Misuse of the linear mixed model when evaluating risk factors of cognitive decline.American journal of epidemiology, 174(9):1077–1088, 2011

    C´ ecile Proust-Lima, Jean-Fran¸ cois Dartigues, and H´ el` ene Jacqmin-Gadda. Misuse of the linear mixed model when evaluating risk factors of cognitive decline.American journal of epidemiology, 174(9):1077–1088, 2011

    2011

  10. [18]

    Instantiated mixed effects modeling of alzheimer’s disease markers.NeuroImage, 142:113–125, 2016

    Ricardo Guerrero, Alexander Schmidt-Richberg, Christian Ledig, Tong Tong, Robin Wolz, Daniel Rueckert, Alzheimer’s Disease Neuroimaging Initiative (ADNI, et al. Instantiated mixed effects modeling of alzheimer’s disease markers.NeuroImage, 142:113–125, 2016

    2016

  11. [19]

    An improved model for disease progression in patients from the alzheimer’s disease neuroimaging initiative.The Journal of Clinical Pharmacology, 52(5):629–644, 2012

    Mahesh N Samtani, Michael Farnum, Victor Lobanov, Eric Yang, Nandini Raghavan, Allitia DiBernardo, Vaibhav Narayan, and Alzheimer’s Disease Neuroimaging Initiative. An improved model for disease progression in patients from the alzheimer’s disease neuroimaging initiative.The J...

    2012

  12. [20]

    Dynamic predictions and prospective accuracy in joint models for longitudinal and time-to-event data.Biometrics, 67(3):819–829, 2011

    Dimitris Rizopoulos. Dynamic predictions and prospective accuracy in joint models for longitudinal and time-to-event data.Biometrics, 67(3):819–829, 2011

    2011

  13. [21]

    A state-space approach for longitudinal outcomes: An application to neuropsychological outcomes.Statistical Methods in Medical Research, 31 (3):520–533, 2022

    Alicia S Chua and Yorghos Tripodis. A state-space approach for longitudinal outcomes: An application to neuropsychological outcomes.Statistical Methods in Medical Research, 31 (3):520–533, 2022

    2022

  14. [22]

    Forecasting the prevalence of preclinical and clinical alzheimer’s disease in the united states.Alzheimer’s & Dementia, 14(2):121–129, 2018

    Ron Brookmeyer, Nada Abdalla, Claudia H Kawas, and Mar ´ ıa M Corrada. Forecasting the prevalence of preclinical and clinical alzheimer’s disease in the united states.Alzheimer’s & Dementia, 14(2):121–129, 2018

    2018

  15. [23]

    Mills, Allan Lawrie, David G

    Mark Cauchi, Andrew R. Mills, Allan Lawrie, David G. Kiely, and Visakan Kadirka- manathan. Individualized survival predictions using state space model with longitudinal and survival data.Journal of The Royal Society Interface, 21(216):20230682, 07 2024. ISSN 1742-

    2024

  16. [24]

    Forecasting individual progression trajectories in alzheimer’s disease

    Etienne Maheux, Igor Koval, Juliette Ortholand, Colin Birkenbihl, Damiano Archetti, Vincent Bouteloup, St´ ephane Epelbaum, Carole Dufouil, Martin Hofmann-Apitius, and Stanley Durrleman. Forecasting individual progression trajectories in alzheimer’s disease. Nature Communicati...

    2023

  17. [25]

    Reflections on dynamic prediction of alzheimer’s disease: advancements in modeling longitudinal outcomes and time-to-event data.BMC Medical Research Methodology, 25, 2025

    Durong Chen, Meiling Zhang, Hongjuan Han, Yalu Wen, and Hongmei Yu. Reflections on dynamic prediction of alzheimer’s disease: advancements in modeling longitudinal outcomes and time-to-event data.BMC Medical Research Methodology, 25, 2025. doi: 10.1186/s12874-025-02618-x

    2025 doi

  18. [26]

    Applying deep learning to predicting dementia and mild cognitive impairment

    Daniel Stamate, Richard Smith, Ruslan Tsygancov, Rostislav Vorobev, John Langham, Daniel Stahl, and David Reeves. Applying deep learning to predicting dementia and mild cognitive impairment. InIFIP international conference on artificial intelligence applications and innovation...

    2020

  19. [27]

    Prediction of progression from mild cognitive impairment to alzheimer’s disease with longitudinal and multimodal data.Frontiers in dementia, 2: 1271680, 2023

    Huitong Ding, Biqi Wang, Alexander P Hamel, Mark Melkonyan, Ting F A Ang, Rhoda Au, and Honghuang Lin. Prediction of progression from mild cognitive impairment to alzheimer’s disease with longitudinal and multimodal data.Frontiers in dementia, 2: 1271680, 2023

    2023

  20. [28]

    Modeling alzheimer’s disease progression using deep recurrent neural networks

    Minh Nguyen, Nanbo Sun, Daniel C Alexander, Jiashi Feng, and BT Thomas Yeo. Modeling alzheimer’s disease progression using deep recurrent neural networks. In2018 International Workshop on Pattern Recognition in Neuroimaging (PRNI), pages 1–4. IEEE, 2018

    2018

  21. [29]

    Longitudinal methods for alzheimer’s cognitive status prediction with deep learning.Alzheimer’s & Dementia, 21 (9):e70488, 2025

    Houjun Liu, Alyssa Mae Weakley, Hiroko H Dodge, and Xin Liu. Longitudinal methods for alzheimer’s cognitive status prediction with deep learning.Alzheimer’s & Dementia, 21 (9):e70488, 2025

    2025

  22. [30]

    Deep learning in alzheimer’s disease: diagnostic classification and prognostic prediction using neuroimaging data.Frontiers in aging neuroscience, 11:220, 2019

    Taeho Jo, Kwangsik Nho, and Andrew J Saykin. Deep learning in alzheimer’s disease: diagnostic classification and prognostic prediction using neuroimaging data.Frontiers in aging neuroscience, 11:220, 2019

    2019

  23. [31]

    Transparency of machine-learning in healthcare: The gdpr & european health law.Computer Law & Security Review, 43:105611, 2021

    Miranda Mourby, Katharina ´O Cathaoir, and Catherine Bjerre Collin. Transparency of machine-learning in healthcare: The gdpr & european health law.Computer Law & Security Review, 43:105611, 2021

    2021

  24. [32]

    Machine learning approaches for speech-based alzheimer’s detection: a comprehensive survey.Computers, 14(2):36, 2025

    Ahmed Sharafeldeen, Justin Keowen, and Ahmed Shaffie. Machine learning approaches for speech-based alzheimer’s detection: a comprehensive survey.Computers, 14(2):36, 2025

    2025

  25. [33]

    Evaluation of speech-based digital biomarkers: review and recommenda- tions.Digital biomarkers, 4(3):99–108, 2020

    Jessica Robin, John E Harrison, Liam D Kaufman, Frank Rudzicz, William Simpson, and Maria Yancheva. Evaluation of speech-based digital biomarkers: review and recommenda- tions.Digital biomarkers, 4(3):99–108, 2020

    2020

  26. [34]

    Resting state eeg biomarkers of cognitive decline associated with alzheimer’s disease and mild cognitive impairment.PloS one, 16(2):e0244180, 2021

    Amir H Meghdadi, Marija Stevanovi´ c Kari´ c, Marissa McConnell, Greg Rupp, Christian Richard, Joanne Hamilton, David Salat, and Chris Berka. Resting state eeg biomarkers of cognitive decline associated with alzheimer’s disease and mild cognitive impairment.PloS one, 16(2):e02...

    2021

  27. [35]

    Neuroimaging advances regarding subjective cognitive decline in preclinical alzheimer’s disease.Molecular Neurodegeneration, 15(1):55, 2020

    Xiaoqi Wang, Weijie Huang, Li Su, Yue Xing, Frank Jessen, Yu Sun, Ni Shu, and Ying Han. Neuroimaging advances regarding subjective cognitive decline in preclinical alzheimer’s disease.Molecular Neurodegeneration, 15(1):55, 2020. 24

    2020

  28. [36]

    Imaging biomarkers associated with cognitive decline: a review.Biological psychiatry, 77(8):685–692, 2015

    Jonathan McConathy and Yvette I Sheline. Imaging biomarkers associated with cognitive decline: a review.Biological psychiatry, 77(8):685–692, 2015

    2015

  29. [37]

    Schein, Alexandrin Popescul, Lyle H

    Andrew I. Schein, Alexandrin Popescul, Lyle H. Ungar, and David M. Pennock. Methods and metrics for cold-start recommendations. InProceedings of the 25th Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’02, page 253–260, Ne...

    2002

  30. [38]

    A systematic literature review of solutions for cold start problem.International Journal of System Assurance Engineering and Management, 15:2818–2852, 2024

    Neetu Singh and Sandeep Kumar Singh. A systematic literature review of solutions for cold start problem.International Journal of System Assurance Engineering and Management, 15:2818–2852, 2024. doi: 10.1007/s13198-024-02359-y. URL https://doi.org/10.1007/ s13198-024-02359-y

    2024 doi

  31. [39]

    Sachs, Gordon J

    Bonnie C. Sachs, Gordon J. Chelune, Stephen R. Rapp, Ashley M. Couto, James J. Willard, Jeff D. Williamson, Kaycee M. Sink, Laura H. Coker, Sarah A. Gaussoin, Tanya R. Gure, Alan J. Lerner, Linda O. Nichols, Carolyn H. Still, Virginia G. Wadley, and Nicholas M. Pajewski. Robus...

    2022

  32. [40]

    Ferri, Mariella Guerra, Yueqin Huang, K

    Ana Luisa Sosa, Emiliano Albanese, Martin Prince, Daisy Acosta, Cleusa P. Ferri, Mariella Guerra, Yueqin Huang, K. S. Jacob, Juan Llibre de Rodriguez, Aquiles Salas, Fang Yang, Ciro Gaona, A. T. Joteeshwaran, Guillermina Rodriguez, Gabriela Rojas de la Torre, Joseph D. William...

    2009 doi

  33. [41]

    Fjell, Linda McEvoy, Dominic Holland, Anders M

    Anders M. Fjell, Linda McEvoy, Dominic Holland, Anders M. Dale, and Kristine B. Walhovd. What is normal in normal aging? effects of aging, amyloid and alzheimer’s disease on the cerebral cortex and the hippocampus.Progress in Neurobiology, 117:20– 40, 2014. ISSN 0301-0082. doi...

    2014 doi

  34. [42]

    Harada, Marissa C

    Caroline N. Harada, Marissa C. Natelson Love, and Kristen L. Triebel. Normal cognitive aging.Clinics in Geriatric Medicine, 29(4):737–752, 2013. doi: 10.1016/j.cger.2013.07.002

    2013 doi

  35. [43]

    Intelligible models for classification and regression

    Yin Lou, Rich Caruana, and Johannes Gehrke. Intelligible models for classification and regression. InProceedings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’12, page 150–158, New York, NY, USA, 2012. Association for Computing Ma...

    2012

  36. [44]

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nature machine intelligence, 1(5):206–215, 2019

    Cynthia Rudin. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead.Nature machine intelligence, 1(5):206–215, 2019

    2019

  37. [45]

    Interpretml: A unified framework for machine learning interpretability.arXiv preprint arXiv:1909.09223, 2019

    Harsha Nori, Samuel Jenkins, Paul Koch, and Rich Caruana. Interpretml: A unified framework for machine learning interpretability.arXiv preprint arXiv:1909.09223, 2019. 25

    1909 arXiv

  38. [46]

    Intelligible models for healthcare: Predicting pneumonia risk and hospital 30-day readmis- sion

    Rich Caruana, Yin Lou, Johannes Gehrke, Paul Koch, Marc Sturm, and Noemie Elhadad. Intelligible models for healthcare: Predicting pneumonia risk and hospital 30-day readmis- sion. InProceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Min...

    2015

  39. [47]

    Interpretable machine learning for precision cognitive aging.Frontiers in Computational Neuroscience, 19:1560064, 2025

    Abdoul Jalil Djiberou Mahamadou, Emma A Rodrigues, Vasily Vakorin, Violaine Antoine, and Sylvain Moreno. Interpretable machine learning for precision cognitive aging.Frontiers in Computational Neuroscience, 19:1560064, 2025

    2025

  40. [48]

    Rand hrs longitudinal file 2022 (v1), May 2025

    RAND Center for the Study of Aging. Rand hrs longitudinal file 2022 (v1), May 2025. Pro- duced with funding from the National Institute on Aging, grant numbers NIA U01AG009740 and NIA R01AG073289, and the Social Security Administration. Santa Monica, CA

    2022

  41. [49]

    Health and retirement study, rand hrs longitudinal file 2022 (v1) public use dataset, 2025

    Health and Retirement Study. Health and retirement study, rand hrs longitudinal file 2022 (v1) public use dataset, 2025. Produced and distributed by the University of Michigan with funding from the National Institute on Aging, grant number NIA U01AG009740. Ann Arbor, MI

    2022

  42. [50]

    URL https://hrsdata.isr.umich

    Delia Bugliari, Nancy Campbell, Chris Chan, Orla Hayden, Jessica Hayes, Michael Hurd, Angelyn Karabatakis, Regan Main, Joshua Mallett, Colleen McCullough, Erik Meijer, Michael Moldoff, Philip Pantoja, Susann Rohwedder, and Patricia St.Clair.RAND HRS Longitudinal File 2022 (V1)...

    2022

  43. [51]

    Fisher, and A

    Mary Beth Ofstedal, Gwenith G. Fisher, and A. Regula Herzog.Documentation of Cognitive Functioning Measures in the Health and Retirement Study. Ann Arbor, MI, 2005

    2005

  44. [52]

    Elias, Alexa Beiser, Philip A

    Merrill F. Elias, Alexa Beiser, Philip A. Wolf, Rhoda Au, Rebecca F. White, and Ralph B. D’Agostino. The preclinical phase of alzheimer disease: A 22-year prospective study of the framingham cohort.Archives of Neurology, 57(6):808–813, 2000. doi: 10.1001/archneur.57. 6.808

    doi

  45. [53]

    Laukka, and Brent J

    Lars B¨ ackman, Sari Jones, Anna-Karin Berger, Erika J. Laukka, and Brent J. Small. Cognitive impairment in preclinical alzheimer’s disease: A meta-analysis.Neuropsychology, 19(4):520–531, 2005. doi: 10.1037/0894-4105.19.4.520

    2005 doi

  46. [54]

    Langa, David R

    Kenneth M. Langa, David R. Weir, Mohammed U. Kabeto, and Amanda Sonnega.Langa- Weir Classification of Cognitive Function (1995 Onward): Data Description. Ann Arbor, MI,

    1995

  47. [55]

    Langa, David R

    Kenneth M. Langa, David R. Weir, Mohammed U. Kabeto, and Amanda Son- nega.Langa-Weir Classification of Cognitive Function (1995–2022): Data Descrip- tion. Ann Arbor, MI, 2025. URL https://hrsdata.isr.umich.edu/data-products/ langa-weir-classification-cognitive-function-1995-2022

    1995

  48. [56]

    R. C. Petersen, P. S. Aisen, L. A. Beckett, M. C. Donohue, A. C. Gamst, D. J. Harvey, C. R. Jack, W. J. Jagust, L. M. Shaw, A. W. Toga, J. Q. Trojanowski, and M. W. Weiner. 26 Alzheimer’s Disease Neuroimaging Initiative (ADNI): Clinical Characterization.Neurology, 74(3):201–20...

    2010 doi

  49. [57]

    Alzheimer’s Disease Neuroimaging Initiative (ADNI), 2024

    Alzheimer’s Disease Neuroimaging Initiative. Alzheimer’s Disease Neuroimaging Initiative (ADNI), 2024. URLhttps://adni.loni.usc.edu/

    2024

  50. [58]

    Andr´ e Rey.Rey Auditory-Verbal Learning Test, 1983

    1983

  51. [59]

    Hammers, Kevin Duff, Kelsey R

    Dustin B. Hammers, Kevin Duff, Kelsey R. Suhrie, Matthew J. Suhrie, Andrew M. Dixon, and Stephanie M. Porter. Assessing and validating reliable change across adni protocols. Journal of Clinical and Experimental Neuropsychology, 44(2):85–102, 2022. doi: 10.1080/ 13803395.2022.2082386

    2022

  52. [60]

    Webb, Yiyi Ou, Tatiana Wilkinson, Andrew M

    Katherine L. Webb, Yiyi Ou, Tatiana Wilkinson, Andrew M. Kiselica, Peter Passmore, Valentina Escott-Price, Bernadette McGuinness, Alan J. Thomas, Paul C. Donaghy, and John-Paul Taylor. Test-retest reliability and minimal detectable change of four cognitive tests in community-d...

  53. [61]

    Hurd, Paco Martorell, Adeline Delavande, Kathleen J

    Michael D. Hurd, Paco Martorell, Adeline Delavande, Kathleen J. Mullen, and Kenneth M. Langa. Monetary costs of dementia in the united states.New England Journal of Medicine, 368(14):1326–1334, 2013. doi: 10.1056/NEJMsa1204629. 27 Supplementary Materials Supplementary Figures ...

    2013 doi

  54. [65]

    Hard negatives were defined as individuals who never had a visit with a worse diagnostic/classification state than their baseline state

    Hard constraint: specificity floor.Configurations were first filtered according to specificity among hard negatives. Hard negatives were defined as individuals who never had a visit with a worse diagnostic/classification state than their baseline state. Only configurations wit...

  55. [66]

    Positive cases not flagged before transition contributed zero reward

    Primary objective: early detection reward.Among configurations satisfying the specificity constraint, we maximized a saturating early-detection reward: R= 1 npos X i∈P 1(ℓi >0) [1−exp(−kℓ i)],(S31) where P is the set of validation-set positives, npos = |P|, ℓi is the lead time...

  56. [67]

    Secondary objective: F2 score.Among configurations tied on the early-detection reward, we maximized the F2 score, a recall-weighted harmonic mean of precision and recall: F2 = 5·precision·recall 4·precision + recall .(S32) Configurations withinϵ F2 of the maximumF 2 score were...

  57. [68]

    Tie-break: visit-level false-positive burden.Among the remaining configurations, the selected configuration was the one with the lowest visit-level false-positive burden among hard negatives, defined as the fraction of hard-negative visits that were flagged. 2.4.3 Age-Stratifi...

    2022

  58. [2020]

    URL https://hrsdata.isr.umich.edu/sites/default/files/documentation/ data-descriptions/Data_Description_Langa_Weir_Classifications2016.pdf

  59. [2022]

    doi: 10.3233/JAD-215564

    doi

  60. [5689]

    URLhttps://doi.org/10.1098/rsif.2023.0682

    doi: 10.1098/rsif.2023.0682. URLhttps://doi.org/10.1098/rsif.2023.0682. 23

    2023

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.