Neurological and psychiatric diseases are fundamentally circuit diseases — conditions where the normal patterns of neural activity are disrupted by cell death, synaptic dysfunction, genetic variants, or structural changes. Understanding them computationally requires connecting molecular pathology to circuit dynamics to clinical phenotype.
This chapter surveys major neurological conditions through the lens of computational neuroscience: what goes wrong biologically, what the computational signature looks like, and how data-driven approaches are changing diagnosis and treatment.
Epilepsy: Runaway Synchrony
Epilepsy is characterized by recurrent seizures — episodes of abnormal, excessive, or synchronous neural activity. The brain transitions from normal heterogeneous firing to pathological hypersynchrony.
The E/I Imbalance Framework
Normal brain function depends on a tight balance between excitation (E) and inhibition (I). Epilepsy represents an E>I imbalance:
- Loss of GABAergic interneurons (reduces inhibition)
- GABA-A receptor mutations reducing inhibitory conductance
- Gain-of-function mutations in voltage-gated Na⁺ channels (increases excitability)
- Loss of K⁺ channel function (reduces repolarization)
Once a focus of hypersynchronous activity is established, it can propagate to recruit neighboring circuits — the "kindling" phenomenon.
Computational Signatures
EEG during seizures shows characteristic patterns:
- Focal onset: high-frequency oscillations (HFOs) at 80–250 Hz in the seizure focus (ripples and fast ripples) — detectable before the clinical seizure
- Seizure propagation: rhythmic spike-and-wave complexes spreading across channels
- Absence seizures: 3 Hz generalized spike-and-wave
Seizure prediction: can we predict a seizure before it happens? EEG shows preictal changes 10–60 minutes before seizure onset — changes in synchrony, connectivity, and spectral composition. ML classifiers trained on these features can predict seizures in some patients with ~80% sensitivity. Implantable devices (RNS System, NeuroPace) continuously monitor EEG and deliver targeted stimulation to abort developing seizures.
Seizure focus localization: for drug-resistant epilepsy, surgical resection of the seizure focus can be curative. Localizing the focus requires intracranial EEG with dense coverage; computational methods for source localization and connectivity analysis improve surgical planning.
Parkinson's Disease: Dopamine, Oscillations, and Movement
Parkinson's disease (PD) is the second most common neurodegenerative disease (~1% of people >60). Cardinal motor symptoms: tremor (at rest), rigidity, bradykinesia (slowness of movement), and postural instability.
Pathophysiology
PD is caused by progressive degeneration of dopaminergic neurons in the substantia nigra pars compacta (SNc) that project to the striatum. The dopamine signal modulates basal ganglia circuits; its loss disrupts the normal balance of direct pathway (movement facilitation) and indirect pathway (movement suppression) activity.
Alpha-synuclein aggregation (Lewy bodies) is the pathological hallmark. The protein misfolds and forms toxic oligomers and fibrils that spread through the brain in a prion-like manner over years before clinical symptoms emerge.
The Beta Oscillation Signature
A key computational finding: dopamine depletion in PD causes excessive synchronization in the beta frequency band (13–30 Hz) in the basal ganglia and motor cortex. This pathological beta synchrony correlates with motor symptom severity:
- Higher beta power → worse bradykinesia
- L-DOPA treatment reduces beta power → symptom improvement
- Deep brain stimulation (DBS) of the subthalamic nucleus (STN) reduces beta synchrony → motor improvement
Beta oscillations in PD are measurable from local field potentials recorded through DBS electrodes — enabling adaptive DBS (closed-loop stimulation triggered when beta power exceeds a threshold), which provides better symptom control with less stimulation.
Computational Models of Basal Ganglia
The basal ganglia circuitry has been extensively modeled:
- Rate models: model firing rates of each nucleus; reproduce the "direct/indirect pathway" framework
- Spiking network models: reproduce beta oscillations from the interaction of STN and GPe neurons
- Reinforcement learning models: map dopamine signals to TD prediction errors; explain action selection and reward learning
These models make predictions about circuit interventions and have guided DBS target selection.
Alzheimer's Disease: Protein Aggregation and Network Failure
Alzheimer's disease (AD) is the most common cause of dementia. Progressive memory loss, then broader cognitive decline, correlates with synaptic loss and neurodegeneration.
Molecular Pathology
Two defining pathological features:
- Amyloid plaques: extracellular deposits of Aβ peptides (cleavage products of APP). The amyloid cascade hypothesis proposes Aβ aggregation as the primary trigger.
- Neurofibrillary tangles: intraneuronal aggregates of hyperphosphorylated tau protein; tau tangles correlate better with cognitive decline than amyloid burden.
Current AD drugs targeting Aβ (lecanemab, donanemab) reduce amyloid load and modestly slow cognitive decline — the first disease-modifying therapies, though clinical benefits are modest.
Network Pathology
AD follows a predictable anatomical progression (Braak stages for tau; amyloid spreads differently):
- Tau begins in the entorhinal cortex → hippocampus → association cortex → primary cortex
- The hippocampus and memory circuits are preferentially affected early → episodic memory loss as the presenting symptom
EEG biomarkers of AD:
- Slowing of EEG (increased theta/delta, decreased alpha/beta)
- Reduced EEG complexity (approximate entropy, sample entropy)
- Disrupted gamma oscillations (possibly due to loss of parvalbumin interneurons)
- Altered functional connectivity
fMRI biomarkers: the default mode network (DMN) shows abnormal connectivity in MCI and AD. DMN regions overlap substantially with regions of amyloid accumulation — suggesting pathological processes specifically target the most metabolically active network.
Computational Approaches in AD
PET imaging analysis: amyloid PET (florbetapir, PiB) and tau PET (flortaucipir) quantify protein aggregates in vivo. Deep learning models that classify PET scans can detect AD years before clinical diagnosis.
Genomics: GWAS has identified ~80 AD risk loci. APOE ε4 is the strongest genetic risk factor (odds ratio ~3–4 for one copy; ~12 for two copies). Multi-polygenic risk scores combining all variants predict disease probability.
Digital biomarkers: speech analysis, gait analysis, and digital cognitive testing show changes years before diagnosis — non-invasive, scalable monitoring.
Multiple Sclerosis: White Matter and Conduction
Multiple sclerosis (MS) is an autoimmune disease in which myelin is attacked by autoreactive T cells, disrupting conduction in white matter tracts. The brain relies on myelination for fast, reliable axonal conduction; demyelination causes conduction failure, slowing, or block.
Computational Impact
The primary computational consequence of demyelination:
- Conduction velocity decreases from ~120 m/s (myelinated) to ~0.5–2 m/s (unmyelinated)
- Increased refractory period → can't maintain high firing rates
- Conduction block at higher frequencies → fatigue with sustained activity
Lesion detection and monitoring: MRI T2-FLAIR lesions map white matter damage. Automated lesion segmentation algorithms (SAMSEG, U-Net variants) quantify lesion burden and track progression.
Diffusion tensor imaging (DTI): measures water diffusion anisotropy as a proxy for white matter tract integrity. Reduced FA (fractional anisotropy) in major tracts precedes lesion formation and correlates with disability.
Depression and Psychiatric Conditions: Circuit Dysregulation
Major depressive disorder (MDD) doesn't have a clear structural pathology — it's a circuit dysregulation condition:
- Hyperactivity of the amygdala (exaggerated emotional responses)
- Hypoactivity of the medial prefrontal cortex (reduced emotion regulation)
- Disrupted default mode network
- Reduced hippocampal neurogenesis (reversible with antidepressants)
Computational biomarkers for depression:
- EEG frontal alpha asymmetry: higher left frontal alpha (less left frontal activity) is associated with depression and negative affect
- EEG theta-alpha coherence: frontal theta coherence with alpha oscillations predicts antidepressant response
- fMRI default mode network connectivity: predicts response to CBT vs. medication
Clinical Neural Decoding and Biomarkers
Across all neurological conditions, the same computational challenge appears: extracting clinically meaningful information from neural signals.
Biomarker development: identify features in neural data that correlate with disease state, trajectory, or treatment response. Key considerations:
- Sensitivity and specificity (ROC analysis)
- Reliability across sessions and individuals (test-retest reliability)
- Biological validity (mechanistic plausibility)
- Clinical feasibility (wearable EEG, at-home monitoring)
Digital phenotyping: smartphones and wearables generate continuous behavioral data (speech, gait, sleep, social interaction) that encodes neurological state. Passive monitoring without clinical visits enables longitudinal tracking at scale — currently being validated in Parkinson's, Alzheimer's, and psychiatric conditions.
Closed-loop therapies: the convergence of sensing and stimulation enables therapies that adapt in real-time to neural state. DBS with beta-triggered adaptive stimulation for PD is the most advanced clinical example. Transcranial alternating current stimulation (tACS) at gamma frequencies is being investigated for Alzheimer's (targeting 40 Hz gamma deficits). The field is moving from fixed schedules to biomarker-guided, closed-loop interventions.
For computational practitioners in this field, the core skills are: signal processing of neural time series, machine learning for biomarker development, clinical trial design for validation, and understanding the translational path from research findings to approved clinical tools.