Using a CNN-LSTM Architecture to Classify Chronic Pain Models based on µECoG Recordings from S1 in large animal models
Authors
Atchuthan, Nickolaj Ajay ; Clark, Hjalte Færregård ; Danyar, Mikkel Bjerre
Term
4. term
Publication year
2023
Submitted on
2023-06-01
Pages
107
Abstract
Baggrund og formål: Kroniske smerter rammer omtrent hver femte voksne i Europa, men diagnostiske metoder er ofte utilstrækkelige. Vi undersøgte, om mikro‑elektrokortikografi (µECoG) fra den primære somatosensoriske cortex (S1) hos store forsøgsdyr (grise) kombineret med en dyb læringsmodel (CNN‑LSTM) kan skelne mellem forskellige eksperimentelle smerte‑modeller. Metoder og materialer: Vi anvendte 16 danske landracegrise. Højfrekvent stimulation (HFS) blev brugt til midlertidigt at fremkalde en kronisk‑lignende tilstand og langtidspotentiering (LTP), mens nogle grise gennemgik spared nerve injury (SNI) for at inducere vedvarende kroniske smerter; andre fungerede som kontrol. ECoG‑målinger blev optaget før og efter indgreb, med lavfrekvent stimulation (LFS) anvendt. Signaldata blev omdannet med en kontinuerlig wavelet‑transform til tids‑frekvens‑spektrogrammer, som blev brugt som input til en CNN‑LSTM for at identificere mønstre, der kunne knyttes til LTP‑, SNI‑ og kontrolgrupperne. Resultater: Testnøjagtigheden var 42,8 % for multiklasse‑modellen og 52 % for den binære model. På testdatasættet lavede begge modeller kun korrekte forudsigelser for kontrolklassen. Valideringsnøjagtigheden var højere: 63,9 % (multiklasse) og 84,0 % (binær). Modelfortolkning viste, at de signalområder og mønstre, som modellen tillagde betydning (feature attribution), var forskellige mellem validerings‑ og testdata. Konklusion: CNN‑LSTM‑modellens ydeevne var utilfredsstillende for både multi‑ og binær klassifikation. For at øge generaliserbarheden bør flere forsøgsdyr, især i LTP‑ og SNI‑grupperne, inkluderes, så træningen indfanger større mellem‑individ variation. Det kan forbedre modellens performance og pålidelighed i fremtiden.
Background and aim: Chronic pain affects about one in five adults in Europe, yet current diagnostic methods are often suboptimal. We examined whether micro‑electrocorticography (µECoG) from the primary somatosensory cortex (S1) in large animals (pigs), combined with a deep learning model (CNN‑LSTM), can distinguish between different experimental pain models. Methods and materials: We studied 16 Danish Landrace pigs. High‑frequency stimulation (HFS) was used to transiently induce a chronic‑like state and long‑term potentiation (LTP), while some pigs underwent spared nerve injury (SNI) to create long‑term chronic pain; others served as controls. ECoG recordings were collected before and after interventions, with low‑frequency stimulation (LFS) applied. Signals were transformed using the continuous wavelet transform to produce time‑frequency spectrograms, which were used as input to a CNN‑LSTM designed to separate LTP, SNI, and control groups by identifying characteristic patterns. Results: Test accuracy was 42.8% for the multiclass model and 52% for the binary model. On the test set, both models made correct predictions only for the control class. Validation accuracy was higher: 63.9% (multiclass) and 84.0% (binary). Model interpretability indicated that the locations and patterns of feature attribution differed between validation and test data. Conclusions: The CNN‑LSTM performed unsatisfactorily on both multiclass and binary tasks. To improve generalizability, more animals—particularly in the LTP and SNI groups—should be included so that training captures greater between‑subject variation. This may enhance performance and reliability in future applications.
[This summary has been rewritten with the help of AI based on the project's original abstract]
Keywords
Documents