I build spiking models of excitable cells (neurons, cardiac myocytes and pancreatic cells) and use them to study how the brain and heart talk to each other, to detect disease early from electrophysiological signals, and to simulate networks at human-brain scale on a desktop computer. My next step takes these models from simulation to the patient: biosensors, embedded signal processing and mobile apps that flag anomalies in the brain, heart and pancreas.
My doctoral work proposed the Hybrid Excitable Cell (HEC) model: a computationally light spiking model with an adaptive threshold, neurotransmitter dynamics and adrenergic receptors. Everything else on this page grows from that model, and it sits where neuroscience, cardiology, endocrinology, electronics and AI meet.
A single set of equations that reproduces 14 neurocomputational firing patterns, cardiac action potentials of eight species, and pancreatic cell activity. The adaptive threshold carries spike-timing history, and adrenergic receptors give the bidirectional brain–organ link.
A fully connected network of 1,000 neurons and cardiac myocytes, linked through α- and β-adrenergic neurons, reproduces the heartbeat as collective neural firing. Heart rate variability (RMSSD) is read directly from the simulated electrophysiology.
Storing each neuron's response as one bit, packed 64 to a word, brings a human-brain-size network of 1011 neurons and 1014 synapses down to about 13 GB. The full run took 1 h 23 min on a Core i5 laptop with 4 GB of RAM.
Hybrid spiking neural networks built on HEC neurons flag abnormal patterns in neural, cardiac and pancreatic signals and predict cardiovascular disease, with the energy profile that suits wearable and implantable devices.
The next stage of my work closes the loop between the models and real people. Biosensors capture the brain, heart and glucose signals; embedded electronics clean and digitise them; HEC-based spiking networks read them; and an Android app warns the user and their clinician when something looks wrong. My background in electronics (MSc Electronics, 15+ years teaching electrical and electronics engineering) is what makes this hardware-to-AI chain possible.
Wearable and implantable sensors for the three organs, connected to one acquisition front end.
Real-time filtering and feature extraction on embedded hardware, close to the sensor.
HEC spiking networks and physiological digital twins that learn each patient's normal pattern.
A mobile-health app that shows the signals, raises alerts and shares reports with clinicians.
Pick a firing pattern from Table 3.1 of my thesis, change the stimulus and add correlated noise. The trace is drawn the way the thesis plots it: membrane potential in blue, adaptive threshold dashed, input current in red.
In-browser demo, forward Euler integration as in Algorithm 1. It uses the quadratic Izhikevich-form dynamics and published presets from which the Table 3.1 parameters were tuned; the dashed line plots the adaptive threshold of Eq. 3.5 after each spike.
Electrophysiological recordings of 2,191 neurons from mouse visual cortex, stimulated optogenetically, over 14 trials of 217 time bins. The HEC model beats GLIF, MAT and QIF on spike-timing and network firing rate.
Each cardiac upstroke fires β-adrenergic neurons; each repolarisation fires α-adrenergic neurons. Their summed firing marks the heartbeat, and the beat-to-beat intervals give heart rate variability, as in Chapter 5 and the Neuroinformatics paper.
Neuron parameters follow §5.5: a, b, c = −65 + 15r², d = 8 − 6r², with r uniform on [0, 1].
RMSSD typically runs from 30 to 150 ms depending on age and fitness; the thesis network gave 40 ms, the value of a healthy older adult. Raise vagal tone to see HRV climb. Action-potential shapes are templates traced from thesis Figs 5.21–5.28, not the full ionic model.
A neuron either fires or it doesn't, so one bit is enough to record it. A 1000 × 1000 matrix of 64-bit numbers holds 64 million neurons; repeating it gives 1011. Colours follow Fig. 4.8: regular, fast, tonic and phasic spiking.
4 features × 1000 × 1000 × 64 bits × 16 trials × 25 ranks
= 1.024 × 10¹¹ bits ≈ 12.8 GB
Conventional brain simulation needs petaflop supercomputers and terabytes of storage. Here each neuron is reduced to its binary state, fired (1) or not (0), so one 64-bit number holds 64 neurons and a 1000 × 1000 matrix holds 64 million.
Synaptic current from 1,000 neurons is summed and passed through the model equations. When the membrane potential v crosses the threshold Vth, the neuron's bit is set to 1 (Algorithm 3). The grid is repeated 16 times per spiking class, for 4 classes, then 25 more times: 1011 neurons in all.
For comparison, the largest network reported so far is about 1 billion neurons (1% of the brain), run on 82,944 nodes of Japan's K computer. The bit-stream approach trades biological detail for memory and speed, which is what puts brain scale within reach of a desktop.
A grid of one million HEC neurons starts at a resting potential of −80 mV, each with 1,000 sparse synaptic connections. Every neuron that reaches 30 mV is drawn as a dot at its (i, j, k) position, so the firing pattern can be read directly on the cortex.
Homogeneous runs give all regions the same five features; heterogeneous runs give each region its own mix. The firings form propagating waves that appear and vanish across the surface, much as in Izhikevich's large-scale thalamocortical model. The same pipeline can be scaled up on SpiNNaker or the K computer and compared with EEG, MEG, ECoG and voltage-sensitive-dye recordings.
An extension from the thesis book: adding glucagon and insulin terms lets the same HEC equations describe the islet cells that control blood glucose.
Diabetes is driven by failures of glucose homeostasis, which the brain regulates through the autonomic nervous system. Adding glucagon- and insulin-induced potentials to the HEC equations reproduces the electrical activity of all three islet cell types.
| Parameter | α-cells | β-cells | δ-cells |
|---|---|---|---|
| a | 0.03 | 0.02 | 0.02 |
| b | −2 | −5 | −1 |
| c | 0 | 50 μF | 50 μF |
| d | 100 | 50 | 60 |
| k | 1 | 3 | 0.3 |
| f | 0.1 | 0.05 | 0.2 |
| Vglu | 0 | 35 mV | 35 mV |
| Vins | 0 | 30 mV | 40 mV |
| VCa | 0 | 70 mV | 70 mV |
| Rk | 0 | 5 MΩ | 5 MΩ |
| RCa | 0 | 5 MΩ | 5 MΩ |
| R | 95 MΩ | 95 MΩ | 95 MΩ |
Integrated with the Euler method, V(k+1) = V(k) + hV′(k). Published in the Arabian Journal for Science and Engineering (2024).
After the thesis, the models were written as digital circuits. One brain chip and three organ chips share one architecture, talk to each other over three wires and report to a phone over Bluetooth. All four are verified designs; none has been fabricated yet.
Press a button, or the Chips link in the menu, to start the chips here. They run the same integer models as the Verilog designs.
One arithmetic circuit serves all cells in turn, one cell per slot of 8 clocks. An organ chip holds 1000 cells; at 8 MHz one step of the whole network takes 1 ms, so it runs in real time.
The chip measures beat interval, beat-to-beat change and action potential duration from the neuron populations alone, and raises seven rhythm flags.
In built-in test conditions it flags tachycardia (137 beats/min), bradycardia (50), an irregular rhythm, a conduction block and a prolonged action potential. With the brain chip connected, the returned drive changes the heart rate: the loop is closed.
These are conditions of the model, not tests on patient recordings.
Given a glucose value from a sensor, the network of islet cells produces its own insulin and glucagon levels. The glucose read back from the cells differs from the value given by 8 mg/dL on average (40–295 mg/dL).
When β-cells are lost or respond weakly, or α-cells do not respond, the read-back no longer follows the glucose, and the chip flags it.
The glucose value comes from a sensor: the companion project GlucoSense is a wearable glucose-monitor application with its own sensor-chip design.
The chip counts the working nephrons and the load on each. Losing 25 % or 50 % of them leaves total filtration at 99–100 %, but the chip raises “nephrons missing” and “hyperfiltration” together: an early-loss pattern.
This model is new. It is not part of the thesis and has not been compared with kidney data; whether such a pattern can be measured in a real kidney is an open question.
A chip that listens to an organ is useful only if its report reaches the person who needs it. The phone is the natural display: it already has a Bluetooth radio and a screen, and the applications are web pages that need no installation.
The applications also run the chips' own models, bit-identical to the Verilog, so they work before any hardware exists. That is how they have been tested; the Bluetooth path has not been tried with a real chip.
These are expectations that follow from the designs. Each has to be demonstrated with fabricated chips and with data from subjects.
The same architecture is planned for further organs (lung, gut, liver, muscle, retina). Next steps: fabrication, and validation of each model against recorded data.
First author on all five. Use Copy citation to paste a reference.
The complete thesis, redesigned for reading: a full-colour layout, opening questions and key takeaways for every chapter, and an appendix with the MATLAB code behind each figure.
Dr. Asif Mehmood · PhD Computer Science · UET Taxila · 2025, chapter 6 extended 2026 · 139 pages · PDF 5.7 MB
One hybrid excitable cell model, followed from ion channels to a network of 100 billion neurons. It covers neurons, cardiac myocytes of eight species and pancreatic cells, a cardio-neural network that predicts heart rate variability, and, in the extended chapter 6, the NeuroBit64 family of chips for brain, heart, pancreas and kidney.
Artificial Neuron Model for Large-Scale Spiking Brain Neural Network, PhD thesis, UET Taxila, 2025. Select a figure to enlarge it.
Thesis: Artificial Neuron Model for Large-Scale Spiking Brain Neural Network. Approved July 2025, conferred September 2025. Supervisor: Dr Muhammad Javed Iqbal.
Research Centre for Modeling and Simulation, Islamabad. CGPA 3.55/4.00. Thesis: Electrical Modeling of Excitable Cells (grade A): neurons and cardiac myocytes as equivalent electronic circuits.
COMWAVE Institute of Sciences and IT, Haripur (Approved Education Centre of SUIT Peshawar). Bachelor-level courses, final-year project supervision, curricula and lab manuals under OBE.
Islamabad.
Mansehra. First Division.
Computational neuroscience, spiking or neuromorphic AI, and biomedical modeling. Referee details are available on request.
