Biomedical Imaging Group Rotterdam · YouTube · 4:47
This 120-second excerpt is an English thesis-talk segment on using grey-matter brain age prediction as a dementia risk biomarker. It first motivates the topic with images contrasting normal aging and disease-related brain change, then defines the key quantity . The speaker cites prior work suggesting this gap can distinguish case and control groups, states a longitudinal goal for dementia risk, and specifies that a deep-learning model is trained on Rotterdam Study MRI data with 5337 subjects and 159 incident dementia cases. This 120-second lecture segment presents results from a deep learning model that predicts brain age from MRI images and defines the gap as predicted brain age minus chronological age. The speaker first evaluates model performance with a scatter plot, reporting MAE = years and ICC = 0.97. He then links the gap to incident dementia using a Cox proportional hazards model, obtaining HR = 1.11 (95% CI 1.06-1.15), and reinforces this with Kaplan-Meier curves stratified by gap quintiles, showing higher dementia-free probability for lower gap groups. Finally, he addresses interpretability by applying Grad-CAM attention maps, highlighting the amygdala and hippocampus as especially influential regions for brain age prediction. The clip closes with a conclusion slide stating that the gap is a significant biomarker associated with dementia risk. This segment presents the conclusions of a thesis on using deep learning for grey matter age prediction. The speaker highlights two main findings: first, the gap between predicted brain age and chronological age is a significant biomarker for dementia risk, where a larger gap correlates with a higher probability of developing the disease. Second, the deep learning approach enables the visualization of specific brain regions important for aging, providing a tool for future neurological research. The segment concludes with academic credits and a thank you message.
Use the learning inspector for key ideas and moments, or open the reading tabs for the complete notes.
Generated from the video's visuals and explanation; not verbatim speech.
The clip opens with event branding and then the presentation title, identifying the topic as grey matter age prediction as a biomarker for risk of dementia.
The first substantive slide motivates the problem visually. Four axial brain images are shown: a normal 25-year-old brain, a normal 75-year-old brain, a 75-year-old brain with mild cognitive impairment, and a 75-year-old brain with Alzheimer's disease. The speaker explains that the brain changes continuously with age and points to hippocampal shrinkage as an example, then adds that neurodegenerative disease and dementia also alter brain structure.
The presentation then shifts from anatomical motivation to a formal pipeline. A flowchart shows a brain image entering a model and producing , and the slide defines the central quantity by . The speaker says recent studies have found that this gap can distinguish case from control groups, which motivates treating it as a potential biomarker for neurodegenerative disease and dementia.
Next, the slide states the study goal: a longitudinal study of the gap in relation to risk of dementia. This marks a transition from cross-sectional prior evidence to the authors' intended temporal analysis.
Finally, the abstract model box is replaced by a blue “Deep Learning” box with a neural-network schematic and the words “Feature extraction + Classification.” The speaker specifies that deep learning is used to build a model taking 3D brain images to predict age. At the same time, the dataset is introduced: Rotterdam Study data from 1990 to the present, with 5337 subjects and 159 incident dementia cases, all used to train, validate, and test the model.
The segment opens on a Methods slide that lays out the core quantity used throughout the talk. A flow diagram maps a brain image through a Deep Learning block to an output labeled age^{}, and beneath it the slide writes . This establishes the central derived feature: a signed difference between model-predicted brain age and actual chronological age.
The slide then changes to Results, where model quality is shown visually and numerically. A scatter plot titled Chronological vs predicted brain age places chronological age on the x-axis and predicted age on the y-axis, both in years. The speaker explains that each red dot is one subject and that the dashed diagonal line corresponds to perfect prediction, i.e. . Below the plot, the reported summaries are years and , so the argument is that the model tracks true age closely on average while retaining a nontrivial error spread.
A third bullet appears under the performance metrics, shifting from prediction accuracy to clinical association. The speaker says that, using the model together with participant data on incident dementia, they assess the association of the gap with dementia incidence via a Cox proportional hazards model. The slide reports with . Because the confidence interval lies entirely above 1, the intended inference is that larger gap values correspond to a higher hazard of developing dementia.
To make the survival relationship more concrete, a Kaplan-Meier panel is added on the right. Its y-axis is Dementia-free probability and its x-axis is Follow-up time in years, with five curves labeled by quintiles of the gap. The visual ordering matters: the lowest quintile curve stays highest, and the highest quintile curve drops fastest. The speaker interprets this as lower gap values being associated with a greater probability of remaining free of dementia over time, providing a graphical complement to the Cox hazard ratio.
The talk then pivots from statistical association to model interpretability. The speaker explicitly notes that interpreting the gap is challenging and asks what it is based on. Because the predictor is a deep learning model, he says they can use visualization methods to trace the result back to the original image. This motivates the next slide rather than introducing a new formula.
A plain brain image is replaced by a slide titled Grad-CAM attention map showing three brain slices with overlaid heatmaps and a color bar running from Low to High. The speaker states that these maps reveal which regions were particularly important for brain age prediction, naming the amygdala and hippocampus. Mathematically, this step does not redefine the gap; instead, it supplies an anatomical explanation of where the trained model placed predictive weight when generating .
The final slide condenses the whole sequence into one conclusion: the gap between predicted brain age and chronological age is presented as a significant biomarker associated with dementia risk. In the logic of the clip, the chain is definition of gap, validation of prediction quality, survival association via Cox and Kaplan-Meier analyses, and finally interpretability via Grad-CAM.
The presentation concludes by defining the brain age gap—the difference between predicted brain age and chronological age—as a critical biomarker. The speaker notes that a higher gap value directly correlates with an increased probability of developing dementia later in life.
Furthermore, the deep learning methodology proved valuable beyond mere prediction. It allowed for the visualization of specific brain regions associated with the aging process, offering a targeted approach for future research into the relationship between brain anatomy and dementia.
The speaker wraps up by stating that this research framework contributes to a deeper understanding of the human brain, followed by the display of academic credits and a final thank you slide.
The core quantity in the excerpt is the difference between a model-predicted brain age and a person's chronological age. The slide presents this as a simple subtraction formula and the speaker describes it verbally as the gap between predicted brain age and actual age.
The speaker cites prior studies reporting that the gap can distinguish case groups from control groups. On that basis, the excerpt frames the gap as a candidate biomarker for neurodegenerative diseases and dementia, and then states a longitudinal goal of studying its utility for dementia risk.
The methodological box on the slide is specialized from a generic “model” to “Deep Learning.” The narration says an artificial intelligence technique called deep learning is used to build a model that takes 3D brain images and predicts age, while the visual adds the internal description “Feature extraction + Classification.”
Before introducing the formula, the talk uses four brain images to separate two sources of structural variation: normal aging across different ages, and disease-related change at the same age of 75 years. This visual contrast supports the idea that brain structure contains age-relevant information that can also be distorted by neurodegeneration.
The empirical basis stated in this clip is the Rotterdam Study, with data spanning 1990 to the present. The slide lists 5337 subjects and 159 incident dementia cases, and the speaker says the data are used to train, validate, and test the model.
The video defines the key feature as the signed difference between predicted brain age and chronological age. A Methods diagram shows a brain image entering a Deep Learning model and producing age^{}, after which the slide writes the equation for the gap. This quantity is distinct from aggregate error metrics because it is subject-specific and can be positive or negative.
Performance is assessed by plotting predicted brain age against chronological age for all subjects. The dashed diagonal marks perfect prediction. The slide reports mean absolute error years and intraclass correlation coefficient 0.97, supporting the claim that predictions are close on average and highly reproducible, though not exact.
The gap is then used as the exposure in a Cox proportional hazards model for incident dementia. The reported hazard ratio is 1.11 with 95% confidence interval 1.06-1.15. Since the interval is entirely above 1, the speaker interprets increasing gap values as associated with a higher chance of dementia incidence.
To visualize the survival pattern, subjects are split into quintiles of the gap and dementia-free probability is plotted over follow-up time. Lower gap quintiles remain higher on the graph, meaning they have a greater probability of staying free of dementia across the observed period. This complements the Cox result with a nonparametric graphical summary.
Because the predictor is a neural network, the talk uses Grad-CAM attention maps to trace the prediction back to image space. The displayed slices use a Low-to-High color scale, and the speaker identifies the amygdala and hippocampus as particularly important regions for brain age prediction. This card explains what the model relied on visually, not a separate statistical test.
A frequent misunderstanding is to conflate the signed gap with the mean absolute error. In this video, the gap is a per-subject feature used in survival analysis, whereas MAE is a nonnegative summary of average prediction discrepancy. They are related to the same two age quantities but play different analytical roles.
The difference between a model-predicted brain age and a patient's actual chronological age serves as a biomarker for dementia. A larger gap indicates a higher risk of developing the condition.
Deep learning models can be used to visualize and identify specific brain regions that are crucial to the aging process, aiding in targeted neurological research.
Explore conditions, steps and evidence. Supplementary explanations are labeled separately from content shown in the video.
The slide shows the term in the equation .
The speaker refers to “predicted brain age” and later says the model predicts age from 3D brain images.
The age predicted by the model from a brain image.
A scalar age value associated with an individual brain scan; units are not stated on the slide.
The slide shows the term in the equation .
The speaker calls this quantity “the person's actual chronological age.”
The person's actual chronological age.
A scalar age value for the same individual used with the predicted brain age; units are not stated on the slide.
The slide defines by the equation .
The speaker describes “a gap between predicted brain age and the person's actual chronological age.”
The difference between predicted brain age and chronological age.
A signed scalar obtained by subtraction; the clip does not state whether absolute values or thresholds are used.
The flowchart labels the input box as “brain image.”
The speaker says the model “takes in 3D brain images.”
The slide text says “brain image,” while the audio specifies “3D brain images”; the exact modality is not written on the slide.
brain image
The input data fed to the model for age prediction.
Brain imaging data; the audio identifies it as 3D brain images, and later the speaker mentions MRI scans from the Rotterdam Study.
The central box in the flowchart is labeled “model.”
At about 01:30 the generic “model” box is replaced by a blue box labeled “Deep Learning” with a neural-network schematic and the text “Feature extraction + Classification.”
The speaker says they used “an artificial intelligence technique called deep learning to build a model.”
model
The computational system that maps a brain image to a predicted age.
In this clip the model is specifically identified as a deep-learning model.
The slide lists “Number of subjects 5337.”
The speaker says the Rotterdam Study gives access to “more than 5000 participants.”
Number of subjects
The size of the Rotterdam Study dataset used in the analysis.
Count of participants; the slide value is 5337.
The slide lists “Incident dementia +159.”
The speaker says there is “a set of 159 participants who developed dementia later on.”
The slide uses the notation “+159,” but the clip does not define the plus sign beyond indicating the count of incident dementia cases.
Incident dementia
The number of participants who developed dementia later on.
Count of incident dementia cases; the slide value is 159.
Slide text shows the equation age^predicted - age^chronological = gap.
The same slide labels the output of the deep learning block as age^predicted.
age^{}
Predicted brain age obtained from a deep learning model applied to a brain image.
Years; scalar output of the regression model.
Slide text shows the equation age^predicted - age^chronological = gap.
Scatter plot x-axis is labeled Chronological age (years).
age^{}
Chronological age of the subject.
Years; observed scalar covariate.
Slide text defines gap by age^predicted - age^chronological = gap.
Speaker later refers to the association of the gap with incident dementia.
gap
Difference between predicted brain age and chronological age.
Years; derived scalar feature used in survival analysis.
Slide bullet reads Mean absolute error years.
Scatter plot annotation reads MAE = 4.45.
Speaker says the mean error was around four and a half years.
Mean absolute error of predicted brain age relative to chronological age.
Years; nonnegative summary statistic.
Slide bullet reads Intraclass correlation coefficient 0.97.
Speaker mentions a high reproducibility score.
Intraclass correlation coefficient quantifying reproducibility of the predictions.
Dimensionless; reported value 0.97.
The slide explicitly writes .
The speaker introduces “a gap between predicted brain age and the person's actual chronological age.”
A flowchart shows “brain image” entering “model,” producing , with the subtraction formula underneath.
The clip defines the brain-age gap as the difference between the age predicted from a brain image and the person's chronological age. The formula is shown directly on the slide and is reinforced by the spoken description.
A predicted brain age must be available from a model applied to a brain image.
A chronological age for the same individual must be known.
The speaker says, “For our study we used an artificial intelligence technique called deep learning to build a model that takes in 3D brain images to predict its age.”
The generic “model” box is replaced by a blue “Deep Learning” box containing a neural-network diagram and the text “Feature extraction + Classification.”
The flowchart input is labeled “brain image” and the output is labeled .
The exact network architecture is not specified in this clip.
The study uses a deep-learning model as the mapping from brain images to predicted age. In the animation, the abstract model box is concretized as a neural network performing feature extraction and classification/regression-style prediction.
Input is a brain image; the audio specifies 3D brain images.
The model outputs a predicted age used in the gap calculation.
The speaker says recent studies have shown that the gap “is able to distinguish case from control groups, indicating that this gap may serve as a biomarker for neurodegenerative diseases and dementia.”
The slide bullet reads “Gap between predicted brain age and chronological age.”
References on the slide include titles about predicting brain-age, reliable and heritable biomarkers, and overlap with common brain disorders.
The clip states prior evidence and motivation; it does not present new statistical results within this excerpt.
Within this excerpt, the brain-age gap is presented as a candidate biomarker because previous work reported that it can distinguish case and control groups in neurodegenerative disease contexts. The clip frames the current study as extending this idea longitudinally.
The gap is computed from predicted brain age and chronological age.
The biomarker claim in this clip is motivated by prior literature rather than demonstrated by new results in the excerpt.
Slide shows age^predicted - age^chronological = gap.
Flow diagram maps brain image through Deep Learning to age^predicted.
Speaker transitions from model building to showing performance.
The video defines the gap as the difference between the model-predicted brain age and the subject's chronological age. The accompanying flow diagram indicates that the predicted brain age comes from a deep learning pipeline operating on a brain image.
Uses a trained deep learning model to obtain age^{}.
Both ages are measured in years.
The gap is a derived scalar feature for downstream analysis.
Slide bullets read Mean absolute error years and Intraclass correlation coefficient 0.97.
Scatter plot annotation reads MAE = 4.45.
Speaker says the mean error was around four and a half years and notes a high reproducibility score.
The results slide reports two quantitative summaries of model performance: mean absolute error and intraclass correlation coefficient. The first measures average prediction discrepancy in years, while the second summarizes reproducibility or agreement strength.
Computed from predicted brain age versus chronological age.
MAE is expressed in years.
ICC is dimensionless and reported as 0.97.
Slide bullet reads Cox regression (hazard ratio of 1.11; 95% confidence interval of 1.06-1.15).
Speaker says they assess the association of the gap with incident dementia using a Cox proportional hazards model.
The video applies a Cox proportional hazards model to test whether the brain-age gap is associated with incident dementia. The reported hazard ratio above one indicates that larger gap values correspond to higher dementia risk.
Outcome is incident dementia.
Exposure is the gap variable.
Model is Cox proportional hazards.
Interpretation requires HR > 1 to indicate increased hazard.
Kaplan-Meier plot titled Kaplan-Meier - Gap shows five curves by quintile.
Speaker says the quintiles of the gap values illustrate the difference in dementia probability between lower and higher gap values.
The presentation visualizes survival differences by dividing subjects into quintiles of the gap and plotting dementia-free probability over follow-up time. Lower gap quintiles remain higher on the plot, indicating greater probability of staying dementia-free.
Subjects are grouped into quintiles by gap value.
Vertical axis is dementia-free probability.
Horizontal axis is follow-up time in years.
Curves are compared across quintile groups.
Slide title reads Grad-CAM attention map with three brain slices and a High-Low color bar.
Speaker says visualization techniques trace results back to the original image and retrieve attention maps.
Speaker states that the amygdala and hippocampus were particularly important for brain age prediction.
To address interpretability of the deep learning model, the video uses Grad-CAM attention maps overlaid on brain slices. The heatmaps indicate which anatomical regions most influenced the predicted brain age, with the amygdala and hippocampus highlighted as especially important.
Requires a trained deep learning model and corresponding input images.
Attention intensity is encoded by a color scale from Low to High.
Used here for post hoc interpretation rather than prediction accuracy.
The speaker states that the gap is a significant biomarker associated with a risk for dementia, and higher gap values were associated with a high probability for developing dementia later on.
Slide text reads: 'The gap between predicted brain age and chronological age is a significant biomarker associated with a risk for dementia.'
The difference between a model's predicted brain age and a person's actual chronological age serves as a biomarker. A larger positive gap indicates a higher probability of developing dementia in the future.
Requires a predictive model for brain age
Applicable to assessing dementia risk
The speaker notes that the deep learning approach allowed visualization of brain regions important for aging, which can be used for future research.
Slide text reads: 'The deep learning approach allowed visualization of brain regions which were important for aging.'
Deep learning models can be utilized not only for predicting outcomes but also for visualizing specific brain regions that are critical to the aging process, guiding further neurological research.
Uses a trained deep learning model
Focuses on interpretability and feature visualization
The speaker says, “Recent studies have shown that a gap between predicted brain age and the person's actual chronological age is able to distinguish case from control groups...”
The slide lists supporting references from 2017 and 2018 on brain-age prediction and related biomarker findings.
The specific case and control groups are not named in the spoken excerpt.
Recent studies have shown that the gap between predicted brain age and chronological age can distinguish case groups from control groups.
Predicted brain age is obtained from brain images.
Chronological age is known for each participant.
The comparison is made across case and control groups in prior studies.
Existential claim about findings in recent studies; the clip does not formalize the quantification.
The speaker continues, “...indicating that this gap may serve as a biomarker for neurodegenerative diseases and dementia.”
The word “may” indicates a proposed interpretation rather than a proven theorem.
The brain-age gap may serve as a biomarker for neurodegenerative diseases and dementia.
The previously observed ability of the gap to distinguish case from control groups is taken as motivating evidence.
Modal claim (“may serve”) rather than a universal statement.
The speaker says, “As a next step, we want to take a longitudinal approach in brain aging over time. Thus we aim to investigate the utility of such a gap as a risk biomarker, for instance, dementia.”
The slide adds “Our goal: Longitudinal study of the gap; Risk of dementia.”
The study aims to use a longitudinal approach to investigate the utility of the brain-age gap as a risk biomarker for dementia.
The gap is computed from predicted brain age and chronological age.
Longitudinal data over time are available for the studied cohort.
Statement of research aim rather than a proved result.
Speaker says the red dots represent all the subjects in the dataset.
Scatter plot contains many red points comparing predicted and chronological age.
Each red dot in the scatter plot corresponds to one subject, with coordinates given by chronological age and predicted brain age.
The plot is the model-performance scatter plot shown on the Results slide.
For every displayed red dot in the scatter plot.
Speaker says the dashed line represents a perfect prediction.
A dashed diagonal line runs through the scatter plot.
The dashed diagonal line in the scatter plot represents the case where predicted brain age equals chronological age.
Axes are chronological age and predicted age in the same units.
For points lying on the dashed line.
Speaker says the model gave a hazard ratio greater than one, indicating that an increasing gap is associated with a higher chance for the incidence of dementia.
Slide reports hazard ratio of 1.11 with 95% CI 1.06-1.15.
Because the Cox model yields HR = , increasing gap values are associated with a higher chance of incident dementia.
The fitted model is a Cox proportional hazards model.
The exposure is the gap variable.
The outcome is incident dementia.
For increases in the gap within the fitted model.
Speaker says lower quintiles or low gap values had a higher probability of being free of dementia.
Kaplan-Meier plot shows the lowest quintile curve above the others over follow-up time.
Across the plotted follow-up period, subjects in lower gap quintiles show a higher probability of remaining dementia-free than those in higher quintiles.
Subjects are partitioned into quintiles by gap value.
The plotted quantity is dementia-free probability over time.
For the quintile groups shown in the Kaplan-Meier plot.
Speaker says these show that the amygdala and hippocampus were particularly important for brain age prediction.
Grad-CAM slices display strong warm-colored activations in medial temporal regions.
The exact anatomical localization is conveyed visually and verbally, but fine-grained voxel-level boundaries are not numerically specified in the clip.
The Grad-CAM attention maps indicate that the amygdala and hippocampus are particularly important regions for the model's brain age prediction.
A trained deep learning model is available.
Grad-CAM is applied to the original brain images.
For the displayed attention maps and the cited model.
The slide shows a left-to-right pipeline: “brain image” → “model” → , with the equation below.
Later the “model” box is replaced by “Deep Learning,” making the computational step explicit.
The speaker explains that deep learning is used to build a model that takes 3D brain images to predict age.
The clip does not show the internal mathematical derivation of the deep-learning model itself.
The flowchart presents the model as a mapping from an input brain image to a predicted age.
Directly shown in the slide diagram and supported by the narration about using a model to predict age from brain images.
Once the predicted age is available, the slide defines the gap by subtracting chronological age.
Explicit formula on the slide.
The abstract model box is later specialized to a deep-learning model with a neural-network illustration.
Animation on the slide replaces “model” with “Deep Learning,” and the speaker names deep learning as the technique used.
The clip’s methodological chain is: use a deep-learning model on a brain image to obtain , then compute for downstream biomarker analysis.
The first methods slide shows four axial brain images labeled Normal 25 years, Normal 75 years, Mild cognitive impairment 75 years, and Alzheimer's disease 75 years.
The speaker says the brain continuously changes with age, using hippocampal shrinkage as an example, then notes that neurodegenerative diseases also affect brain structures.
The speaker then asks how this affects information retrieved about a person's age and cites recent studies showing the gap distinguishes case from control groups.
The argument is motivational rather than a formal proof; the clip does not derive the biomarker claim mathematically from the anatomy images.
The talk begins from the premise that the brain changes as a person ages.
Stated directly by the speaker and illustrated by comparing 25-year-old and 75-year-old normal brains.
Because aging leaves information in brain structure, the speaker says the brain can be used to estimate a person's age.
Explicit spoken inference from the preceding anatomical example.
The speaker adds that dementia and neurodegenerative disease also alter brain structures.
Explicit spoken statement accompanying the comparison of mild cognitive impairment and Alzheimer's disease images.
These disease-related deviations are operationalized through the difference between predicted brain age and chronological age.
Formula shown on the next slide after the motivational question.
Prior findings that the gap separates groups are used to motivate its possible role as a biomarker.
Spoken summary of recent studies and supporting reference list on the slide.
The clip builds an intuitive rationale: age and disease both alter brain structure, so a model-based mismatch between predicted brain age and chronological age may carry biomarker information.
Speaker explains that using the model and participant data, they assess the association of the gap with incident dementia using a Cox proportional hazards model.
Slide reports hazard ratio of 1.11; 95% confidence interval of 1.06-1.15.
Start from the previously defined gap feature.
Given on the Methods slide at the beginning of the clip.
Use the gap as the explanatory variable in a survival model for incident dementia.
Stated by the speaker when introducing the analysis.
Report the fitted hazard ratio and its confidence interval.
Shown on the Results slide.
Interpret a hazard ratio above one as increased hazard with increasing gap.
Explicitly stated by the speaker.
The video connects the defined gap variable to a statistically significant increase in dementia risk through the reported Cox hazard ratio.
Kaplan-Meier plot shows five curves labeled by quintile.
Speaker says lower quintiles or low gap values had a higher probability of being free of dementia.
Group the cohort by ordered gap values into five quintiles.
Stated by the speaker and reflected in the legend.
For each quintile q, plot the estimated dementia-free probability over follow-up time t.
Axes and legend of the Kaplan-Meier figure define this quantity.
The curves are ordered from top to bottom by increasing gap quintile.
Visible in the plot and summarized verbally by the speaker.
The Kaplan-Meier visualization supports the claim that smaller gap values correspond to a higher probability of remaining dementia-free over time.
The slide shows two normal brains labeled “25 years” and “75 years,” with visible differences in highlighted internal structures.
The speaker says, “As an example, the hippocampus is shown over here, where aging reflects into a shrinkage of the structure.”
The highlighted region is identified in speech as the hippocampus, but the slide itself does not print the word “hippocampus.”
Illustrate how normal aging changes brain structure in a way that could support age estimation.
A 25-year-old normal brain image.
A 75-year-old normal brain image.
The speaker identifies the highlighted structure as the hippocampus.
Show that aging produces visible structural change in the brain.
The slide places a younger and an older normal brain side by side to make age-related change visually explicit.
Direct visual comparison on the slide.
The speaker interprets the visual difference as shrinkage associated with aging.
Audio explanation tied to the displayed images.
Normal aging is illustrated by structural shrinkage in the highlighted brain region between the 25-year-old and 75-year-old examples.
The conclusion is supported by the side-by-side images and the speaker’s explicit identification of shrinkage with aging.
The slide compares a Normal 75-year-old brain with Mild cognitive impairment 75 years and Alzheimer's disease 75 years.
The speaker says, “However, neurodegenerative diseases and dementia also affect brain structures.”
The clip does not quantify the visible differences; it only presents them illustratively.
Show that disease can alter brain structure even when chronological age is held constant.
A Normal 75-year-old brain image.
A Mild cognitive impairment 75-year-old brain image.
An Alzheimer's disease 75-year-old brain image.
Demonstrate that neurodegenerative disease affects brain structure independently of the age label.
All three compared brains carry the same age label, isolating disease status as the changing factor.
Visible labels on the slide.
The slide juxtaposes different diagnostic categories at the same age.
Direct visual layout of the three rightmost brain images.
The speaker uses this comparison to argue that neurodegenerative disease also affects brain structures.
Audio statement paired with the displayed images.
At the same chronological age of 75 years, the normal, mild cognitive impairment, and Alzheimer's disease brains are shown differently, illustrating disease-related structural effects.
Supported by the same-age labels and the speaker’s explicit statement that neurodegenerative diseases and dementia also affect brain structures.
The slide lists “Rotterdam Study data (1990 – present): Number of subjects 5337; Incident dementia +159.”
The speaker says the data come from the Rotterdam Study, with MRIs of more than 5000 participants used to train, validate, and test the model, plus 159 participants who developed dementia later on.
The exact split of the 5337 subjects into training, validation, and test sets is not given in this clip.
Specify the data source and cohort sizes used to build and evaluate the brain-age model.
Rotterdam Study data spanning 1990 to present.
5337 total subjects.
159 incident dementia cases.
Establish the empirical basis for the longitudinal biomarker investigation.
The slide names the cohort and its temporal span.
Visible caption on the slide.
The total sample size is listed explicitly.
Visible caption on the slide and paraphrased in speech as “more than 5000 participants.”
The subset developing dementia later on is listed explicitly.
Visible caption on the slide and repeated in the audio.
The speaker states that the dataset is used across the full model-development pipeline.
Audio statement.
The study uses Rotterdam Study data with 5337 subjects total and 159 incident dementia cases, applying the data to train, validate, and test the deep-learning model.
The numbers are directly readable on the slide and match the spoken summary.
Scatter plot titled Chronological vs predicted brain age with axes in years and annotation MAE = 4.45.
Speaker explains the axes, red dots, dashed line, and summary error metrics.
Interpret the displayed scatter plot of chronological age versus predicted brain age and the accompanying summary statistics.
x-axis: Chronological age (years)
y-axis: Predicted age (years)
red dots: subjects in the dataset
dashed line: perfect prediction
MAE = years
ICC = 0.97
Explain what the plot and metrics say about prediction quality.
Each subject contributes one point whose horizontal coordinate is chronological age and vertical coordinate is predicted brain age.
Explicitly described by the speaker.
Points on the dashed diagonal correspond to perfect prediction.
Speaker states the dashed line represents a perfect prediction.
The average absolute discrepancy between predicted and chronological age is about 4.45 years.
Shown on the slide and paraphrased in speech.
The reproducibility score is high, indicating strong agreement.
Slide bullet plus speaker's mention of a high reproducibility score.
The scatter plot shows predicted brain age closely tracking chronological age along the identity line, with reported MAE years and ICC 0.97 indicating good but imperfect performance.
Visual clustering near the dashed diagonal plus the explicit numerical metrics on the slide support the interpretation.
Kaplan-Meier - Gap plot with five quintile curves and y-axis Dementia-free probability (-), x-axis Follow-up time (years).
Speaker interprets the separation between lower and higher gap quintiles.
Use the Kaplan-Meier plot to compare dementia outcomes across quintiles of the gap.
Five quintile groups by gap value
Vertical axis: dementia-free probability
Horizontal axis: follow-up time in years
Determine which gap groups remain dementia-free more often over time.
Locate the topmost curve labeled lowest quintile.
Legend and visible ordering in the plot.
Observe that the highest quintile curve lies below the lower quintile curves over follow-up.
Direct visual comparison of the plotted lines.
At comparable follow-up times, lower gap groups have higher dementia-free probability.
Speaker explicitly states this interpretation.
Subjects in lower gap quintiles have a higher probability of remaining dementia-free across follow-up time than those in higher quintiles.
Consistent with both the plotted curve ordering and the speaker's verbal interpretation.
Three brain slices under the title Grad-CAM attention map with a High-Low color bar.
Speaker says the amygdala and hippocampus were particularly important for brain age prediction.
The precise slice coordinates are not given in the clip.
Interpret the Grad-CAM attention maps shown for the trained model.
Title: Grad-CAM attention map
Color bar from Low to High
Three overlaid brain slices
Identify which regions the model attends to most strongly.
Warmer colors indicate higher attention according to the color bar.
Diagram label High/Low.
The most intense regions appear around structures identified in speech as amygdala and hippocampus.
Combination of visual hotspots and speaker's anatomical naming.
Grad-CAM is used to trace the model result back to the input brain image.
Speaker explicitly describes this purpose.
The attention maps highlight the amygdala and hippocampus as particularly important regions for the model's brain age prediction.
Supported by the visible warm-colored activations and the speaker's direct interpretation.
A title card reads “YOUNG MEDICAL DELTA PRESENTS: THESIS AWARDS 2020,” followed by a slide with “Johnny Wang” and “Grey Matter Age Prediction as a Biomarker for Risk of Dementia.”
The speaker introduces himself as John and states the talk title.
The visible name is “Johnny Wang,” while the speaker says “my name is John.”
Title card
Speaker webcam inset
Slide with presenter name and talk title
The screen moves from the awards title card to the presentation title slide.
The talk topic remains grey matter age prediction as a dementia risk biomarker.
This segment establishes the presentation context and topic before any method content appears.
A “Methods” slide shows four axial brain images labeled Normal 25 years, Normal 75 years, Mild cognitive impairment 75 years, and Alzheimer's disease 75 years.
A red laser-pointer dot moves among the brain images as the speaker discusses aging and disease effects.
The exact anatomical boundaries of the highlighted region are not labeled on the slide.
Four axial brain images
Diagnostic labels
Age labels
Red pointer dot
The pointer shifts attention from the younger normal brain to the older normal brain and then to the disease examples at age 75.
All four images remain visible on the same slide during this interval.
The age labels and diagnostic labels stay fixed while the pointer moves.
The visual sequence supports the spoken contrast between normal aging and disease-related structural change.
The previous brain images disappear and a new layout builds on the same “Methods” slide.
A flowchart appears with “brain image” → “model” → and the formula .
Three references appear beneath the formula, then “Our goal: Longitudinal study of the gap; Risk of dementia” is added.
Brain icon
Model box
Output label
Gap formula
Reference list
Goal bullets
The slide transitions from anatomical examples to an abstract pipeline.
References are added under the formula.
The goal statement appears after the biomarker motivation.
The central relation remains the subtraction formula defining the gap.
This animation converts the earlier qualitative motivation into a formal computational pipeline and research objective.
The generic “model” box is replaced by a larger blue box labeled “Deep Learning” containing a neural-network schematic and the text “Feature extraction + Classification.”
The lower text changes from the goal bullets to “Rotterdam Study data (1990 – present): Number of subjects 5337; Incident dementia +159.”
The neural-network drawing is schematic and does not specify layer counts or architecture details.
Deep Learning box
Neural-network schematic
Dataset caption
Citation line
The abstract model becomes a concrete deep-learning module.
The goal bullets are replaced by dataset information and a citation.
The overall pipeline still runs from brain image to and then to the gap formula.
The visual update identifies the computational method and the empirical cohort used to instantiate the framework.
Slide shows brain image -> Deep Learning -> age^predicted, with equation age^predicted - age^chronological = gap.
Logos TU Delft, Erasmus MC, and BIGGR are visible on the left sidebar.
brain image
Deep Learning box
age^{} label
gap equation
Rotterdam Study data bullet points
Static slide is shown briefly before transition to Results.
The pipeline direction from image to predicted age remains fixed.
The gap formula stays visible during this interval.
This visual establishes the computational setup: a deep learning model maps a brain image to predicted brain age, and the gap is defined relative to chronological age.
Scatter plot titled Chronological vs predicted brain age with red points and dashed diagonal; annotation MAE = 4.45.
Speaker explains axes, red dots, dashed line, and performance metrics.
red scatter points
dashed diagonal line
x-axis Chronological age (years)
y-axis Predicted age (years)
MAE annotation
bullet metrics below plot
A pointer moves across the plot and then toward the metric bullets.
Around 41 s, an additional Cox regression bullet appears below the existing metrics.
The axes and overall point cloud remain unchanged during the explanation.
The dashed identity line remains fixed as the reference for perfect prediction.
The plot visually demonstrates agreement between predicted and chronological age, while the annotations quantify error and reproducibility.
Right side of slide adds Kaplan-Meier - Gap plot with five colored curves and legend by quintile.
Speaker introduces the Kaplan-Meier graph and interprets quintile separation.
five stepwise survival curves
legend with quintile labels
y-axis Dementia-free probability (-)
x-axis Follow-up time (years)
The Kaplan-Meier panel appears on the right side of the Results slide.
Pointer moves among the curves and legend while the speaker compares groups.
All curves start near probability 1 at time 0.
The ordering of curves by quintile remains consistent across the displayed follow-up range.
The animation-like slide change introduces a survival visualization showing that lower gap quintiles maintain higher dementia-free probability over time.
Slide changes from a single grayscale brain image to three brain slices titled Grad-CAM attention map with a High-Low color bar.
Speaker says visualization techniques trace results back to the original image and retrieve attention maps.
single grayscale brain image
three brain slice panels
warm-colored activation overlays
High-Low color bar
citation to Selvaraju et al. (2017)
First a plain brain image is shown.
Then the slide switches to three overlaid attention-map slices.
Pointer moves across the highlighted regions.
The color bar orientation from Low to High remains constant once the Grad-CAM slide appears.
The three-slice layout remains fixed during the explanation.
This visual sequence illustrates how the abstract model output is grounded back into anatomical space, revealing which brain regions drive the prediction.
Final slide titled Conclusions states: The gap between predicted brain age and chronological age is a significant biomarker associated with a risk for dementia.
Conclusions heading
single bullet statement
Slide changes from Grad-CAM maps to a text-only conclusion.
No further mathematical content is added after the conclusion appears.
The closing slide restates the main inferential takeaway: the brain-age gap functions as a significant dementia-risk biomarker.
A presentation slide titled 'Conclusions' displays two bullet points summarizing the research findings, followed by presenter and supervisor credits.
Presentation slide
Text bullets
Presenter video feed
Bullet points appear sequentially on the slide
Credits appear at the bottom of the slide
Slide title remains 'Conclusions'
Presenter remains visible in the top right corner
The visual structure organizes the final takeaways of the research, separating the core scientific findings from the academic attribution.
The slide transitions to a black background with large pink text reading 'THANK YOU.' and a logo for 'medical delta YOUNG THESIS AWARDS 2020'.
Black slide background
Pink text
Award logo
Background changes from white to black
Text changes to 'THANK YOU.'
Presenter video feed remains visible
Signals the formal end of the presentation.
The slide distinguishes from and defines their difference as the gap.
The speaker contrasts “predicted brain age” with “the person's actual chronological age.”
One might treat the model output as if it were the person's true age.
In this clip, is a model-derived quantity from a brain image, while is the actual age; the scientific signal of interest is their difference, the gap.
The speaker says the gap “may serve as a biomarker” and that the study aims to investigate its utility as a risk biomarker.
The slide frames the topic as risk of dementia rather than definitive diagnosis.
One might infer from the word “biomarker” that the gap directly diagnoses dementia.
The excerpt presents the gap as a candidate risk biomarker motivated by prior group-separation findings and by the study goal, not as a confirmed diagnostic criterion within this clip.
The speaker first describes normal aging effects and then explicitly adds, “However, neurodegenerative diseases and dementia also affect brain structures.”
The slide separately shows normal aging at different ages and disease conditions at the same age of 75 years.
One might collapse all brain structural differences into “aging” alone.
The clip distinguishes ordinary age-related change from additional disease-related alteration, and this distinction motivates studying the mismatch between predicted brain age and chronological age.
Slide defines gap as age^predicted - age^chronological, without absolute value.
Later analysis treats increasing gap as associated with higher dementia risk.
One might assume the gap is the same thing as the mean absolute error or always nonnegative.
The video defines gap as a signed difference, age^{} - age^{}, whereas MAE is a separate nonnegative performance metric. The survival analysis concerns the gap itself, not the absolute error.
Speaker explicitly says hazard ratio greater than one indicates that an increasing gap is associated with a higher chance for the incidence of dementia.
A reader could mistakenly interpret HR = 1.11 as protective or negligible.
In the Cox model shown here, HR > 1 means higher gap values are associated with higher incidence of dementia; the reported CI [1.06, 1.15] lies entirely above 1.
Y-axis is labeled Dementia-free probability (-), not incidence rate.
Speaker discusses probability of being free of dementia.
One may read the curves as directly giving number of dementia cases.
The plotted quantity is the estimated probability of remaining dementia-free over follow-up time. Differences between curves reflect survival probabilities by quintile, not unadjusted case counts.
Speaker frames Grad-CAM as giving visual feedback on the detection and tracing results back to the original image.
Heatmaps indicate attention intensity rather than lesion burden or causal pathology.
Bright regions in Grad-CAM could be mistaken for direct proof that those structures cause dementia.
In the video, Grad-CAM is used to explain which image regions the deep learning model relied on for brain age prediction. This is an interpretability tool for the model, not a standalone causal claim about disease mechanism.
The speaker moves from examples of age- and disease-related brain change to the question of what information can be retrieved about a person's age.
The first methods slide shows anatomical comparisons; the next slide formalizes these ideas as a model-based prediction pipeline.
The conceptual need to estimate age from brain structure motivates introducing a model that maps brain images to predicted age, which is then used in the gap definition.
The gap formula requires as one of its terms.
The flowchart explicitly routes the model output into the subtraction formula.
The definition of the gap depends operationally on first obtaining a predicted brain age from the model.
The biomarker discussion directly refers to “this gap” after the formula has been introduced.
The slide bullet names the quantity as the gap between predicted brain age and chronological age before discussing its utility.
The biomarker concept applies the previously defined gap quantity to the problem of distinguishing disease-related groups and assessing dementia risk.
The speaker says, “As a next step, we want to take a longitudinal approach in brain aging over time.”
The slide adds “Our goal: Longitudinal study of the gap; Risk of dementia.”
The stated goal generalizes the cross-sectional biomarker motivation to a longitudinal setting by studying the same gap over time for dementia risk.
The slide replaces the goal bullets with “Rotterdam Study data (1990 – present): Number of subjects 5337; Incident dementia +159.”
The speaker says the data come from the Rotterdam Study and were used to train, validate, and test the model.
The deep-learning method is instantiated on the Rotterdam Study dataset described in the example.
Gap is defined first on the Methods slide; MAE and ICC are reported afterward on the Results slide.
The performance metrics compare predicted brain age with chronological age, so they depend on the prior definition of predicted age and the gap framework introduced at the start.
Speaker says they assess the association of the gap with incident dementia using a Cox proportional hazards model.
Earlier slide defines gap explicitly.
The Cox analysis uses the gap as its explanatory variable, so the definition of gap must precede the survival modeling step.
After discussing the Cox model, the speaker adds, Additionally, we can show the Kaplan-Meier graph.
Both analyses concern dementia outcomes and the gap variable.
The Kaplan-Meier display provides a complementary survival visualization for the same gap-related dementia association summarized numerically by the Cox model.
Methods slide shows Deep Learning producing age^predicted.
Speaker says because we have used a deep learning model, we can apply visualization techniques to trace our results back to the original image and retrieve the attention maps.
Grad-CAM interpretation presupposes the existence of the trained deep learning model that generated the predicted brain age and hence the gap.
Gap is defined as a signed difference; MAE is reported separately as years.
The gap is a subject-specific signed feature used in downstream risk analysis, whereas MAE is an aggregate nonnegative measure of prediction accuracy.
The speaker explicitly links the deep learning approach to the ability to visualize brain regions.
The deep learning method is applied to generate the predicted brain age, which is then used to calculate the gap biomarker.
The slide directly displays .
The speaker says prior studies show the gap distinguishes case from control groups and may serve as a biomarker.
The model box becomes a Deep Learning box with a neural-network schematic.
The speaker says deep learning is used to build a model that takes 3D brain images to predict age.
The slide compares normal aging across ages and disease states at the same age.
The slide lists 5337 subjects and 159 incident dementia cases from the Rotterdam Study.
The slide states “Our goal: Longitudinal study of the gap; Risk of dementia.”
The speaker says they want to take a longitudinal approach in brain aging over time.
Slide explicitly defines gap by equation.
Results slide lists MAE and ICC values.
Scatter plot compares predicted and chronological age.
Cox regression bullet gives HR 1.11 and CI 1.06-1.15.
Speaker interprets HR > 1 as higher dementia chance with increasing gap.
Kaplan-Meier plot shows quintile-separated dementia-free probability curves.
Speaker explains lower quintiles have higher probability of being dementia-free.
Grad-CAM attention map slide with High-Low color bar.
Speaker names amygdala and hippocampus as particularly important.
Signed gap formula and separate MAE metric both appear in the clip.
Covered · Opening title cards and speaker introduction establish the talk topic but contain no separate mathematical definition beyond the title context.
Covered · Methods slide with four brain images and spoken motivation about aging and neurodegenerative disease effects.
Covered · Slide builds the brain-image → model → predicted-age pipeline and defines the gap; speaker cites prior biomarker evidence.
Covered · Goal bullets appear and the speaker states the longitudinal dementia-risk aim.
Covered · Model box becomes Deep Learning; dataset statistics for the Rotterdam Study are shown and narrated.
Covered · Methods slide defines the prediction pipeline and the gap formula.
Covered · Results slide presents the scatter plot and performance metrics MAE and ICC.
Covered · Cox regression bullet appears and the speaker interprets the hazard ratio.
Covered · Kaplan-Meier plot by gap quintile is introduced and interpreted.
Covered · Slide transitions to brain images and Grad-CAM attention maps for interpretability.
Covered · Conclusions slide restates the main takeaway; no new mathematical content beyond the summary claim.
Covered · First conclusion point regarding the biomarker.
Covered · Second conclusion point regarding deep learning visualization.
Covered · Slide displays presenter and supervisor credits; no new mathematical content.
Covered · Thank you slide and end of presentation.
Reviewed English material v1: card/concept_card "Cox regression result for incident dementia" starting at 2:41 — The gap is then used as the exposure in a Cox proportional hazards model for incident dementia. The reported hazard ratio is 1.11 with 95% confidence interval 1.06-1.15. Since the interval is entirely above 1, the speaker interprets increasing gap values as associated with a higher chance of dementia incidence. ,\%\
YouTube7:31Another explanation of Regression modeling.
Regression modeling
YouTube3:54Another explanation of Regression modeling.
Regression modeling
YouTube3:40Another explanation of Regression modeling.
Regression modelingYouTube3:54Next in Mathematical modeling: Regression modeling → Least squares and residual diagnostics.
Least squares and residual diagnostics · View learning sequence
YouTube3:07Next in Mathematical modeling: Regression modeling → Least squares and residual diagnostics.
Least squares and residual diagnostics · View learning sequence
YouTube8:04Next in Mathematical modeling: Regression modeling → Least squares and residual diagnostics.
Least squares and residual diagnostics · View learning sequenceYouTube7:31Visual explanation for Regression modeling, compared with the current lecture.
Regression modelingYouTube3:54Visual explanation for Regression modeling, compared with the current lecture.
Regression modeling
YouTube7:04Application of Regression modeling, complementing the current explanation.
Regression modeling