Proprietary ML Models for Road Traffic Safety
and Presymptomatic Medical Diagnostics

​​

​The Challenge

​

Current diagnostics of Parkinson's disease (PD) rely on clinical motor symptoms (such as resting tremors) which emerge after 60–80% of dopaminergic neurons in the brain have undergone irreversible degeneration. At this stage, the opportunity for neuroprotection has passed, leaving patients confined to purely symptomatic management. While the pharmaceutical industry is actively developing treatments to slow PD progression, the success of these disease-modifying therapies (DMTs) — none of which are yet clinically approved — depends entirely on early intervention before discernible symptoms appear. The economic burden of PD has surpassed €75 billion per year solely in the EU and the US. Broadening the scope, the aggregate impact of prevalent neurodegenerative conditions (PD, AD) is unprecedented, with nearly 75 million individuals affected and over €1.7 trillion in annual social costs.

​

​Scientific Background

​

The seamless integration of visuospatial perception, cognitive faculties and fine motor coordination required to operate a vehicle serves as a definitive benchmark of neurological health. Our algorithms are able to detect disease-specific changes in driving habits that are characteristic of neurodegenerative disorders (as well as chronic non-neurological conditions like diabetes, arthritis and glaucoma) 4–8 years before the onset of clinical symptoms. This methodology captures the functional impact of the first 5–10% of neuronal degeneration, effectively bypassing the limitations of other medical protocols. Off-the-shelf smartphones, utilizing only their built-in IMU and GPS sensors, are sufficient to reach such a diagnostic edge.

 

However, our approach is not built on the futile attempt to teach an ML model how people with PD drive. Such traditional supervised learning is fundamentally flawed, since patients with presymptomatic PD drive "normally" by all conventional standards. Any hypothetical micro-movements at this stage will be physically indistinguishable from the noise caused by a poorly maintained road, worn vehicle components, or even in-seat movements of a passenger. We likewise deliberately ignore high-level metrics such as driving duration, chosen routes, or average speed, as these universal classifiers cannot be treated as serious medical instruments.

 

Instead, we have pioneered and scientifically substantiated a method to extract the "biological truth" of internal energy and attention allocation within the human brain. By integrating raw sensor data with infinite-dimensional modelling of the longitudinal evolution of our proprietary topological constructs, we identify the fundamental neurological mechanisms that supervise the selection of temporary driving tactics. Our system monitors in what way the brain prioritizes its limited resources under the pressure of incipient neurodegeneration — characterizing how compensatory workarounds (such as ongoing cortical remapping) consume the neuronal computing power previously dedicated to the fine-tuning of motion trajectories, acceleration dynamics and real-time response patterns. By detecting this subtle internal shift in "resource cannibalization", we capture the precise moment when the brain's effort to maintain apparent normality modifies its subclinical operational efficiency.

​

Consequently, in the case of Parkinson's disease, the process we identify is not the resting tremor — a symptom which will remain imperceptible for another 5 years due to homeostatic plasticity. We monitor how the patient's brain allocates its resources at the most fundamental level (e.g., controlling lane position, managing speed, and reacting to external stimuli while masking early-stage neurodegenerative deficits through increased axonal sprouting).​

​

​Technological Engine

​

Once the collected raw data (IMU and GPS measurements) is cleaned and enriched via advanced feature engineering, it is processed through our proprietary "analytical neural network" architecture (a purpose-built LSTM with a tailored attention mechanism and a domain-specific cost function). Moving beyond standard supervised learning, we employ our novel adversarial inverse reinforcement learning (AIRL) framework — enhanced by latent behavioural embedding capabilities — to automatically infer the individual driving reward function of each patient. Central to our diagnostic methodology is an original mathematical construct established through foundational scientific research — referred to as the "diagnostic function". This non-linear operator explicitly allows us to identify the functional impact of the first 5–10% of neuronal degeneration by mapping AIRL-derived driving reward functions into a specialized, infinite-dimensional feature space associated with a positive-definite kernel (referred to as the "diagnostic space").

 

By treating driving as an optimization problem, we can mathematically model the underlying logic behind the user's subconscious decisions. The mapping of driving reward functions into the diagnostic space yields what we define as the "driving phenotype" — a dynamic analytical entity that serves as a high-resolution surrogate for the patient's neurological state, and whose continuous evolution reflects the underlying biological reality of the brain. Our system tracks the kinematics of an individual's driving phenotype — its translational drift (the velocity and direction of its movement through the diagnostic space), its rotational dynamics (the specific axes and angular velocities), and the modalities by which its internal structure is reconfigured. This multifaceted alteration allows us to capture the metabolic cost of the brain's attempt to maintain standard performance despite progressing anomaly.​

​

​Differential Diagnosis

​

To distinguish between PD, AD, ALS, Parkinson-plus syndromes, and non-neurological conditions (e.g., glaucoma, diabetes, rheumatoid arthritis, or general fatigue), our system utilizes a Temporal Geometry-Aware Transformer (referred to as the "diagnostic neural network") trained on synchronized EEG, fNIRS, eye tracking (saccadic latency, fixation precision, microsaccade dynamics), sEMG, GSR, HRV, pressure mapping, and telematics data acquired within our proprietary VR driving simulator. Therefore, the long-term evolution of a driving phenotype serves as a unique representation of precisely defined neurophysiological dysfunctions (e.g., those detected during our ML model training within the cortico-subcortical loops, the cerebro-cerebellar pathways, or the prefrontal-striatal circuits).

 

The methodology described above allows us to identify PD as the only incipient physiological state consistent with the entire constellation of observed features. While behavioural anomalies detected on the road — induced by prodromal degeneration of dopaminergic neurons — may appear non-specific or mimic general cognitive decline, our framework decomposes these macroscopic traits into millions of latent parameters. In our infinite-dimensional diagnostic space, the cumulative probability of any non-PD condition (e.g., normal ageing, effects of medication, or multiple system atrophy) generating this particular evolutionary pattern of the driving phenotype is statistically negligible. Therefore, PD remains the only mathematically compatible solution for the observed phenomena.

​​

The synthesis of state-of-the-art neurobiology, pioneering manifold topology, and clinical-grade ground truth from synchronized data fusion has enabled us to conclusively delineate the diverse landscape of evolutionary patterns of driving reward functions. Consequently, our diagnostic verdict constitutes a scientifically substantiated conclusion: we observe a mathematical phenomenon so deeply intertwined with the expected mechanics of dopaminergic decay that, in the light of current medical knowledge, PD remains the only viable explanation for its emergence. Perhaps most remarkably, this architecture has been engineered without prior clinical labelling of the study participants. The aforementioned "PD-specific diagnostic class" has materialized from the vast conjunction of multi-modal data streams. By aggregating subtle telematics and neuroimaging correlations from a diverse cohort, our system has effectively built a hitherto unknown "Parkinsonian profile" from the intersection of individual neurobiological states.​​​

​

​The Results

​

We have observed that specific anomalies in the driving reward function trajectory — identified by our system as related to abnormal brain resource allocation characteristic of prodromal PD development — evolve predictably into the overt disease patterns seen in advanced PD. Explicitly, the same early shifts of driving phenotypes manifest as a nascent form of PD biomarkers that subsequently intensify while the disease progresses into the clinical phase. This longitudinal alignment provides robust empirical validation of our model, confirming that it captures the fundamental neurodegenerative trajectory long before traditional diagnostic thresholds are reached.

 

TCR App Diagnostics exhibits the necessary discriminative precision to differentiate the digital signatures of PD from those characteristic of Parkinson-plus syndromes, such as multiple system atrophy (MSA). This successful topological delineation stems from the analysis of unique disease modalities — including the pace of progression and the probability of attributing observed signals to the specific neuroanatomical footprint of each disorder. Simultaneously, the efficacy of our diagnostic approach does not rely on the manual execution of driving tasks, but on the cognitive and sensorimotor integration required to supervise and interact with the vehicle. Our methodology is entirely independent of driving experience, advanced driver assistance systems (ADAS), mechanical issues (e.g., worn tires), environmental conditions (e.g., heavy rain, road irregularities), random events (e.g., in-seat movements of a passenger) and temporary or permanent vehicle changes. Our novel adversarial inverse reinforcement learning (AIRL) framework treats all of these factors simply as a new baseline, where the underlying mathematical signatures of neurodegeneration remain consistent and inevitably emerge.​​

​

​The Impact

​

TCR App Diagnostics delivers the discriminative precision necessary to differentiate PD from Parkinson-plus syndromes, eliminates the requirement for invasive lumbar punctures and hospital visits, expands the potential diagnostic reach from thousands of patients typically served in clinical settings to millions of individuals monitored in their natural environment, and enables a dynamic view of patient health rather than a single snapshot of their biological state at the time of the procedure. Our approach significantly outperforms both conventional neurological assessments, such as the UPDRS, and advanced molecular imaging like DaT-SPECT. Furthermore, it offers critical advantages that CSF-based biomarker assays (e.g., alpha-synuclein) cannot provide.​

​

The transformative breakthrough of our methodology is anchored in its departure from traditional diagnostic paradigms. Conventional AI approaches in PD detection require massive, longitudinal datasets collected from already diagnosed patients — a process that is not only prohibitively expensive and time-consuming but also prone to significant geographic and ethnic biases. Our system bypasses the aforementioned limitations by focusing on the fundamental allocation of cognitive resources. The AIRL-derived driving reward function provides a mathematical description of how the human brain achieves an optimal driving policy. Since our model establishes an individual baseline in the diagnostic space devised to monitor the evolution of these advanced analytical representations, it eliminates the need for market-specific adjustments or large-scale data collection every time our technology is implemented in a different geographical region. Consequently, TCR App Diagnostics can enter any global market and immediately define a local norm, ensuring objective diagnostic accuracy that is inherently immune to cultural or demographic variances in driving attitudes and ageing traits.

​

​Clinical Pathway

​

Once our mobile application is installed, the user is not obliged to perform any specific actions, undergo manual tests, or interact with the system in any way. Our software is fully compatible with older devices, operates in the background (consuming minimal phone resources), and intelligently filters noise from relevant data streams. It functions as a sophisticated data acquisition tool, passively collecting IMU and GPS measurements during the user's daily driving activities — whether the phone is mounted on the dashboard, placed anywhere within the vehicle, or kept in a pocket. Our framework eliminates the need for any additional hardware or specialized equipment. Since TCR App Diagnostics does not utilize cameras, there is no requirement to orient the smartphone towards the driver's face or the road.

​

Raw, anonymized data is securely transmitted to our cloud platform for regular processing (all complex ML operations are performed exclusively there). After a data collection period spanning 14 days to 12 months — depending on the stage of PD and assuming a benchmark of three hours of driving per week — our system compiles enough longitudinal measurements to generate a comprehensive medical assessment or a conclusive diagnosis (if specific disease indicators are present). The detailed report is then conveyed to the patient's healthcare provider. Our product is designed for prescription-based distribution, guaranteeing that every user remains under the continuous supervision of a qualified healthcare professional. Simultaneously, it is intended to act as an independent diagnostic tool, providing automated results that drive the clinical path.

​​

TCR App Diagnostics bridges the gap between high-dimensional data science and frontline clinical neurology. While our underlying technological engine utilizes complex reinforcement learning and topological analysis, the clinical results are translated into a standardized, intuitive medical report that aligns with established neurological nomenclature.

​

​Beyond Neurodegeneration: Primary Brain Tumours

​

While the current operational focus of our startup is strategically directed towards the prodromal detection of Parkinson's disease — in response to the urgent clinical demand and established regulatory pathways — its underlying mathematical architecture is broadly applicable across a spectrum of chronic conditions. By mapping the AIRL-derived driving reward functions into the infinite-dimensional diagnostic space, our system successfully delineates distinct "diagnostic classes" that represent the physiological impact of both neurodegenerative (PD, AD, ALS, PD-plus syndromes) and non-neurological disorders (e.g., glaucoma, diabetes, rheumatoid arthritis).

​

This extensive library of disease signatures allows for a revolutionary approach to brain oncology: by defining the boundaries of known metabolic and neurological drifts, we can identify incipient brain tumours as high-entropy anomalies that defy all validated disease trajectories. Consequently, our system is able to detect these focal aberrations not by recognizing a singular oncological profile, but by quantifying a radical departure from all documented healthy and abnormal states, offering a critical window for intervention during the earliest stages of neoplastic development (WHO Grade I).

 

Unlike neurodegenerative diseases, which follow structured and predictable pathways, brain tumours exhibit extreme heterogeneity in terms of histological type, growth velocity, and anatomical localization. Therefore, it is mathematically impossible to define an "oncological trajectory" within the diagnostic space. However, TCR App Diagnostics overcomes this limitation. When a patient's driving phenotype reveals a seemingly stochastic divergence that is inconsistent with all settled patterns of ageing, neurodegeneration, or chronic illness, our framework identifies it as a non-specific outlier. While the diagnostic neural network cannot definitively label such an anomaly as a tumour, it can describe a level of topological disruption so extreme that it warrants immediate clinical attention. In such cases, our system acts as a high-precision triage tool, recommending a prioritized MRI referral to investigate the underlying cause.