Memory models

Computational Psychiatry: Using a model of memory to track memory decline

Holly Hake, University of Washington, Seattle [email protected]

Graduate Student Holly Hake

Today

> Groups > Create model > Assess performance > Test diagnostic potential > Connect to brain

We want to model the computational level of cognition to better connect the underlying processes in neural level with what we observe at the knowledge (behavioral) level.

Let’s create a model of memory

To create a memory model, let’s first start by thinking about a problem we would like it to solve. Problem: You have an exam in your Italian course in 10 days on the names of different pasta shapes. You want to create a study paradigm that will ensure you remember the information for the exam. This model would represent long-term, episodic memory.

Let’s create a model of memory

Study time (days) 1 2 3 4 5 6 7 8 9 10

What study paradigm would ENSURE we remember the test material?

Think about the 3 memory research phenomena we learned last week

Let’s create a model of memory

Study time (days)

RECENCY

1 2 3 4 5 6 7 8 9 10

We could study close to the exam. More recent information is more likely to be retrieved.

Let’s create a model of memory

Study time (days)

FREQUENCY & RECENCY

1 2 3 4 5 6 7 8 9 10

We could study close to the exam, AND study every day; high frequency. BUT, we don’t have time for that. We need to ENSURE we will remember this information AND make the most EFFICIENT use of our time.

Let’s create a model of memory

Study time (days)

FREQUENCY & RECENCY

1 2 3 4 5 6 7 8 9 10

We could start by lowering the frequency. However, consider the scenario:

Two students follow this EXACT paradigm. They do everything exactly the same. Student A gets 100%, but student B gets a 75%. What do you think could have been the individual differences that contributed to this?

Think back to what we learned about

memory decay.

Let’s create a model of memory

time

% c

or re

ct

This is an example of how memory decays over time. A classic law of cognition is that forgetting curves can be closely approximated by power functions.

Let’s create a model of memory Each time you encounter a fact, a new trace is created (so we are basing our idea of how memories are encoded using the Multiple Trace Theory) The overall “memory” of that fact is the summed collection of all it’s traces.

Forgetting curve for m

P(m)/P(¬m)

Frequency: cumulative effect of traces Recency: decay over time This forgetting curve now shows that the time you learn a fact for the 4th time, you have a higher odds of retrieving the memory, and the memory itself decay a lot slower.

Forgetting Threshold

= Fusilli

Lo g

od ds

Moment at which m is predicted to be forgotten

Forgetting Threshold (50/50 chance of remembering)

Let’s test student A and B based on when our model predicts they have a 50/50 chance of remembering. (If our guess on individual differences is correct, then student B should reach their forgetting threshold sooner than student A.

Let’s make a prediction

= Fusilli

Lo g

od ds

Moment at which m is predicted to be forgotten

Test when m should be forgotten

?

?

Lo g

od ds

= Fusilli

If m is remembered, reduce RoF ↓

New estimate for when m will be forgotten

Lo g

od ds

= Fusilli

= Fusilli

If m was forgotten, increase RoF ↑

#!#&*

?

Chair /Kiti

Lo g

od ds

New estimate for when m will be forgotten

= Fusilli

Let’s make a model of memory

> This model can now make predictions about an individual’s memory, but also refine those predictions in real time. – This makes the model more individualized!

> Each time you encounter a fact, a new trace j is encoded

> A memory m is the summed collection of the j traces > Each trace j decays over time in a power law

P(m)/P(¬m) = ∑j t( j ) -d

Model of episodic memory based on Multiple Trace Theory

P(m) / P(¬m) = ∑j t( j ) -d

> α = Rate of Forgetting > Accounts for spacing effect > Modify decay rate- make it a function of the previous

activation of the memory > If you know the history of a memory m, its

probability of being forgotten depends on α

d = eA(m) + α

Adding the spacing effect

P(m) / P(¬m) = ∑j t( j ) -d

> Calculates activation 15s in the future > Adapts α after 3 fact repetitions > Multiply response time by 1.5 if answer is incorrect > t0 = fixed amount of time

for the perception and motor processes involved in giving the answer

d = eA(m) + α

Calculating RoF (α)

E(RT) = e−Ax + t0

Summary

> We created a lightweight version of a model that could model the probability of retrieving a particular memory over time.

> Our computational cognitive model that simulates memory encoding (with the multiple traces and with how each memory is a sum of those traces) and passive forgetting (with decay rate) based on established cognitive and biological principles.

> Now that we have an idea of how this process is happening, and we can capture it with a series of eqs, then individual differences could be captured as a series of parameters for these eqs.

Computational Psychiatry: Using a model of memory to track memory decline

Now that we have a memory model that measure and predict an individual’s rate of forgetting, let’s use that to track memory decline!

Tracking memory decline is difficult

Current Methods

Standardized questionnaires

New Method (Online memory game)

Increase overall reliability, convenience, repeatability, and sensitivity

Project

> Measure Rate of Forgetting > Find neural substrates of forgetting

Participants

subjective or mild cognitive impairment (MCI)

age-matched controls

100100

Mild Cognitive Impairment (MCI)

> People whose functional capacity is relatively intact, but who, on objective testing, show cognitive decline in at least one area of neuropsychological functioning

> Why this pop? – High risk for progression to clinical dementia – More time to track

Clinical Evaluation

> Health History > Family Health History > Clinical Dementia Rating (CDR®)

– Six domains of cognitive and functional performance: Memory, Orientation, Judgment & Problem Solving, Community Affairs, Home & Hobbies, and Personal Care.

> Montreal Cognitive Assessment MoCA > Neuropsychological Battery

– Craft Story 21 Recall (Immediate) – Benson Complex Figure Copy – Number Span Test: Forward – Number Span Test: Backward – Category Fluency – Trail Making Test – Craft Story 21 Recall (Delayed) – Benson Complex Figure Recall – Multilingual Naming Test (MINT) – Verbal Fluency: Phonemic Test

> Neurological evaluation

Neuroimaging

> MRI: – Resting state fMRI – Diffusion Tensor Imaging

> PET: – FDG – Tau – Amyloid

Biomarkers

> CSF – Amyloid – Abeta40 – Abeta42 – pTau – Total Tau

Behavior

> Memory Tasks – 52 Weeks – 8 min 1 x week

> Website: https://adrc.slimstampen.nl/#/lessons?tab=my-lessons > User info: [email protected] password

How reliable is the Rate of Forgetting?

Avg Rate of Forgetting (α) for each lesson

Distribution of avg Rate of Forgetting (α) for each lesson

Test-retest reliability of Rate of Forgetting (α)

> r = 0.71

Test-retest reliability of Rate of Forgetting (α) without the first lesson

> Exclude each participants first lesson to account for the initial learning curve

> r = 0.74

Test-retest reliability of Rate of Forgetting (α) by type of stimuli

> Separate by text vs visual stimuli

> TEXT r = 0.79 > VISUAL r = 0.62

Distribution of Rate of Forgetting (α) by type of stimuli

> Separate by text vs visual stimuli

> TEXT α = 0.4 > VISUAL α = 0.4

How diagnostic is the Rate of Forgetting?

Avg Rate of Forgetting (α) for each lesson by clinical status

Distribution of Rate of Forgetting (α) by clinical status

> On avg, Healthy controls have an RoF of 0.37, whereas MCI has an RoF of 0.42

Distribution of Rate of Forgetting (α) by type of stimuli and clinical status

> Separate by text vs visual stimuli AND clinical status

> MCIs forget text stimuli faster than HCs

Aunt

Mother

Holly

There is a 95% chance that you will have forgotten a new fact after:

~ 1-2 mins (Dementia) ~ 11 mins (Aunt) ~ 45 mins (Mother) ~ 2.5 hrs (Holly)

ROC curve for Rate of Forgetting (α)

> Can predict MCI diagnosis with 90% accuracy

Probability of MCI diagnosis by Rate of Forgetting (α)

Takeaways

> A single 8-min online memory game can diagnose Mild Cognitive Impairments with ~80% accuracy

Are there neural correlates of Rate of Forgetting?

Neural data related to Rate of Forgetting (α)

> Predictive connectivity

> Node importance

> mPFC nodes in the DMN correlate most with Rates of Forgetting

Neural data related to Rate of Forgetting (α)

Neural data related to Rate of Forgetting (α)

> Increases in rates of forgetting will mirror the progression of the disease > Rates of forgetting will be correlated with differences in mPFC connectivity > Greater rates of forgetting will be associated with greater white matter

hyperintensities, smaller hippocampal volume, and higher values of global brain atrophy (Rane et al., 2020)

> Disparity between forgetting rates modeled for recognition vs recall paradigms will be associated with frontal white matter hyperintensities

> Concordant forgetting rates for recognition and recall paradigms will be associated with hippocampal and adjacent temporal lobe atrophy

Manipulate Rate of Forgetting (α) through neurostimulation

> Transcranial Direct Current Stimulation (tDCS) was selected as the mode of stimulation because it has successfully been shown to enhance working memory performance (Ke et al., 2019).

> The area of interest was the left hippocampus (LHPC) because of its role in memory encoding.

> Electrodes placed at TP7 (anode) and EX14 (cathode) using the International 10-10 EEG system.

Manipulate Rate of Forgetting (α) through neurostimulation

> Stim Protocol – Active anodal stimulation -

1.5 mA for a duration of twenty minutes

– Sham control – 1.5 mA for a duration of one minute

Manipulate Rate of Forgetting (α) through neurostimulation

> Initial pilot (n = 7) showed no differences in pre and post

> Next step: Target mPFC regions identified in previous experiments