Danny

Inspiration

The inspiration primarily comes from when we spoke to Danny, a Whitechapel local in the East London area. Danny has been struggling with Parkinson's for the past seven years and it hasn't been easy. Based on our limited interaction, we were able to get a brief outlook on what his day-to-day life looks like. Limited caretaker support, infrequent hospital visits and being on multiple types of medication has left him living a less fulfilling life than he could be having. The reason we are building Danny, a personal caretaker for people with terminal illness, is so that him and others like him don't have someone to rely on all the time for simple tasks which have grown to become much more difficult, although this is something they would never admit.

What it does?

We analyse the data of Parkinson's patients from the ; Eventually the aim of this project, is to take real-time data and use it to train a model so that we can eventually predict Parkinson's disease. So the ideation flow would be to use tabular foundation models on the current datasets give and to train the model to identify tremors using the accelerometer connected to the Arduino, after this, the data entered by the caretaker or the patient in terms of medication time-stamps and meal timings and severe tremors will then again be used to train the model creating a flywheel where more and more data is obtained when the model is running and the model is trained on this very data making it more and more accurate and personalized the more time it runs.

How we built it

We created a unified 3 part system connecting Arduino Uno (R4 with Wifi) with a Grove base shield which has an included 3 axis digital accelerometer +-1.5g v1.3 on D8 and is placed on the wrist to detect individual tremor frequency signals. We then collect individual medication timelines (for this prototype, restricted to 3 times a day every 5 hours, 5 times a day every 3 hours or patient's own variable schedule) from either the patient, current caretaker who assists with medication or NHS/Hospital nursing staff who deal with later-stage Parkinson's conditions, as they are unable to operate hand-held mobile devices.

Both are combined to design and implement an end-to-end tabular foundational model architecture that fuses two real-time user signals: 1.⁠ ⁠Medication Timing Log (timestamps, formulation, levodopa equivalent daily dose [LEDD]) 2.⁠ ⁠Wrist Tremor Frequency & Intensity (derived from continuous wearable IMU, accelerometry at 50–100 Hz, aggregated into 15–30 min chunks).

alongside feature sets derived from two clinical reference cohorts:

•⁠ ⁠Dataset A (UK Biobank Proteomics): Baseline neurodegenerative risk priors (NfL_w_score, GFAP_w_score, eGFR, demographics, polygenic risk scores). •⁠ ⁠Dataset B (CUE1+ Clinical Trial / Digital Phenotyping): Fine-Motor Digital Tests (BRAIN_KS, DFT_KS), Wearable Kinematics (PKG_Bradykinesia_Score, PKG_Dyskinesia_Score, wear compliance), and Clinical Motor Subscores (MDS-UPDRS Part III bradykinesia, tremor, and rigidity sub-scores).

The above datasets were then used to reconstruct an "ECB Motor Graph" (Effect-Concentration-Biomarker Graph) that maps: •⁠ ⁠X-axis: Latent / Estimated Striatal Dopamine Concentration (effective central levodopa compartment level). •⁠ ⁠Y-axis: Real-World Medication Timing & Progression Across the Day ( aligned to scheduled dosing intervals and wear windows). •⁠ ⁠Z-plane / Colormap: Observed Motor State (Tremor power spectrum 3–7 Hz, Bradykinesia score, or dyskinesia probability).

The application from this learned manifold formulates a closed-loop personalized recommendation engine to predict optimal, proactive redosing intervals before dopamine depletion causes severe "OFF" state, which in this case would be a return of tremor.

The tremor frequency data will be provided from an accelerometer v1.3 3 axis (+-1.5g) in D8 on an arduino R4 with wifi and a hat. The accelerometer represents the movement of a wristband, which can be processed before putting it in the foundation model.