Neural mass models describe the aggregate behaviour of large, densely interconnected populations of neurons rather than the activity of individual cells. By operating at the population scale, they map naturally onto the rhythms recorded in electroencephalography (EEG) and, by extension, onto disorders of cognition in which those rhythms are disturbed.
Alzheimer’s disease is a natural application. The cholinergic hypothesis holds that reduced synthesis of acetylcholine (ACh) is a central driver of the disease, and the lateral geniculate nucleus (LGN) — the thalamic relay of the visual pathway — is among the most densely innervated targets of brainstem cholinergic projections. Because the brainstem also receives a copy of the sensory signals passing through the thalamus, it sits at the junction of sensory and motor processing. That position implicates the brainstem-to-thalamus cholinergic pathway in the visuomotor impairments seen in the early stages of the disease and makes it a worthwhile target for computational investigation.
This study set out to build a more biologically faithful model of that pathway. It takes an existing neural mass model of the LGN and embeds within it an explicit kinetic model of cholinergic synaptic transmission, thereby imitating the brainstem’s ACh input to the thalamus.
The methodological choice is deliberate. Earlier models often represented synaptic transmission with a generic “alpha function” applied indiscriminately across neurotransmitter types. Replacing that with a receptor-level kinetic framework is both more biologically plausible and computationally tractable. In this framework, the fraction of open postsynaptic ion channels evolves according to forward and reverse reaction rates driven by neurotransmitter concentration. The concentration of transmitter released by the brainstem is itself modelled as a sigmoidal function of brainstem membrane voltage, so presynaptic activity feeds through to postsynaptic current in a graded, physiologically motivated way.
The circuit comprises three interacting cell populations within the LGN: thalamocortical relay cells (TCR), interneurons (IN) and the thalamic reticular nucleus (TRN). Excitatory retinal input reaches TCR and IN via AMPA-mediated synapses; TCR excites TRN, also through AMPA; TRN returns inhibitory GABAergic feedback to TCR; and IN provides feedforward GABAergic inhibition.
Onto this established architecture, the model adds brainstem cholinergic synapses whose sign depends on the receptor type of the target population — excitatory onto the excitatory TCR population and inhibitory onto the inhibitory IN and TRN populations. The resulting system of first-order differential equations is integrated using the fourth- and fifth-order Runge–Kutta–Fehlberg (RKF45) method in MATLAB over a 40-second window at 1 ms resolution. Outputs are averaged across twenty simulation runs to suppress the influence of noisy model input.
Two principal findings emerge. First, introducing cholinergic input produces a significant change in the postsynaptic voltages of all three populations relative to the acetylcholine-free baseline. This confirms that ACh materially shapes the thalamocortical dynamics that give rise to the oscillations observed in EEG and validates the inclusion of an explicit cholinergic pathway rather than an implicit or lumped one.
Second, and more pointedly, the model reveals a sharp sensitivity localised to a single connection. When the strength of the brainstem-to-TRN cholinergic synapse is reduced progressively, the TRN postsynaptic potential falls in step. But when that connection is severed entirely — setting its connectivity parameter to zero to simulate a complete block of ACh to the TRN — the system does not simply continue along the same trend. It undergoes a distinct bifurcation, expressed as a marked hyperpolarisation of the TRN population.
The two remaining populations behave very differently under the same manipulation: the interneurons show only a slight hyperpolarisation, while the excitatory relay cells are essentially unaffected by either graded reduction or full disconnection. The effect is therefore specific to the brainstem-to-TRN link rather than a generic response to cholinergic loss.
The contribution of the work lies in that specificity. By embedding receptor-level cholinergic kinetics into a thalamocortical neural mass model and driving it with RKF45, the study identifies a computationally reproducible signature of cholinergic failure and pins it to a particular pathway. Because the TRN governs sensory gating and the generation of rhythmic thalamocortical activity, an abrupt change of state in this population offers a plausible circuit-level correlate of the visuomotor impairment reported in early Alzheimer’s disease and, potentially, a model-based biomarker for early detection.
The study received both Best Paper of the Session and Best Paper of the Conference at the 7th International Conference on Soft Computing for Problem Solving (SocProS) in 2017.
Several limitations frame the result and point towards future work. The analysis is confined to postsynaptic effects; presynaptic contributions — in particular, anomalies in the release of acetylcholine itself — are not yet included and may act as an additional control parameter for the bifurcation. The GABAB pathway is omitted because the source model’s output is largely insensitive to it, although its role remains to be assessed directly. The parameter values, while consistent with ranges reported in the underlying literature, are in places scaled to remain numerically tractable.
None of these caveats undercuts the central observation. Rather, they mark out the boundaries of a diagnostic framework that later work has extended from detection towards models of intrinsic, astrocyte-mediated repair.