Using Wearable Device Data for Step Measurement on Parkinson's Disease Population

Using Wea able De ice Da a o S ep Measu emen on Pa kinson’s Disease Popula ion Howon Ryu1 1Uni e si y o Cali o nia, San Diego, He be We heim School o Public Heal h and Human Longe i y Science, San Diego, CA, USA Abs ac Pa kinson's disease (PD) is a p og essi e neu odegene a i e diso de wi h a ious mo o symp oms. Home de ec ion and moni o ing o such symp oms p o e o be aluable, as i enables mo e cons an moni o ing a pa ien 's con enience. Wea able de ices equipped wi h ine ial measu emen uni (IMU) senso s a e pa icula ly essen ial in objec i e symp oms p og ession moni o ing a home. Some li e a u es iden i y gai ea u es, which cha ac e ize a pe son's walking o unning mo emen , as impo an p edic o s o de ec ing PD symp oms. Such gai ea u es can be de i ed om IMU signals. In his wo k, we p opose a s ep measu emen me hodology using con olu ional neu al ne wo k a chi ec u e, which is an in eg al s ep in de i ing impo an gai ea u es. Wi h he limi ed accessibili y o such gai ea u es, an open-sou ce s ep-measu emen model ha ansla es aw IMU signals in o gai ea u es would be aluable o esea che s in Pa kinson's disease. We demons a e he use o he p oposed model h ough he Wea Gai -PD da ase . Key Wo ds: digi al heal h, wea able de ice, Pa kinson’s disease, ine ial measu emen uni , ime se ies, con olu ional neu al ne wo k 1. In oduc ion Pa kinson’s disease (PD) is a p og essi e neu odegene a i e diso de cha ac e ized by a ange o mo o symp oms including emo s o change in gai pa e ns such as eezing o gai . Fo PD ea ly de ec ion, con inuous and objec i e moni o ing o such symp oms is c i ical. Home-based de ec ion and moni o ing app oaches a e pa icula ly impo an , as hey enable mo e equen and consis en acking wi hou he need o pa ien s o egula ly isi clinical se ings in pe son. In his ega d, wea able de ices equipped wi h ine ial measu emen uni (IMU) senso s ha e eme ged as aluable ools, se ing as objec i e means o assessing PD p og ession. IMU ypically consis s o accele ome e s and gy oscopes, wi h an occasional inclusion o magne ome e s; IMU acks mo emen dynamics and spa ial o ien a ion in h ee-dimensional space. Speci ically, accele ome e s measu e accele a ion along he X, Y, and Z axes, while gy oscopes cap u e angula eloci y ac oss he same dimensions. Toge he , he IMU signals p o ide a desc ip ion o how an objec o a pe son mo es o e ime. Figu e 1: Example plo o IMU signal, showing accele ome e x, y, and z eadings. IMU signals p esen se e al challenges in analysis. IMU signals a e inhe en ly ime se ies and high-dimensional. De-noising is o en necessa y because o he noise in he inpu s emming om de ice placemen , calib a ion, o senso d i [1-2]. Addi ionally, IMU signals consis o mul iple modali ies o da a, such as accele ome e and gy oscope, which adds o he complex co ela ional s uc u e o he inpu da a whe e he [x, y, z] eadings o one modali y a e highly co ela ed while he eadings ac oss di e en modali ies a e co ela ed o a lesse deg ee. The complexi y o he IMU signals as inpu da a demands ad anced modeling app oaches. One e ec i e way o summa ize IMU da a is h ough gai ea u es, which a e me ics ha cha ac e ize walking o unning pa e ns. Gai ea u es include measu es such as cadence (s eps pe minu e), s ide leng h (dis ance be ween he wo oo s eps o he same side), and s ep leng h (dis ance be ween he wo oo s eps o di e en sides). These me ics a e pa icula ly ele an in PD, whe e changes in walking pa e ns o en e lec disease onse and p og ession. By condensing complex IMU signals in o in e p e able measu es, gai ea u es se e as a b idge be ween aw senso da a and clinically meaning ul p edic ion models. P io li e a u e consis en ly demons a es how using gai ea u es con ibu e o e ec i e and in e p e able PD symp om classi ica ion and p edic ion models (Table 1). Howe e , accessibili y o gai ea u es emains a signi ican challenge. Mos gai me ics a e de i ed using p op ie a y algo i hms embedded wi hin IMU de ices which limi s accessibili y. Fu he mo e, de ice-speci ic algo i hms also limi gene alizabili y ac oss da ase s o di e en de ices, as ea u es may no be compa able. Gi en hese challenges, he e is a need o open-sou ce models ha de i e gai ea u es di ec ly om aw IMU signals. Such models would enable ep oducible and mo e uni e sal gai de i a ion. One p omising app oach in ol es p edic ing oo con ac in o ma ion om he aw IMU signals, as oo con ac p o ides he ounda ion o compu ing a wide ange o gai ea u es. By de ec ing oo con ac in o ma ion, we c ea e a s epping s one o de i ing clinically ele an gai ea u es wi hou elying on p op ie a y algo i hms. In his espec , we p opose a deep lea ning model designed o p edic oo con ac om aw IMU eco dings. Speci ically, we implemen ed a 1D con olu ional neu al ne wo k (CNN) wi h skip connec ions, ained on accele ome e and gy oscope signals. The CNN-based a chi ec u e e ec i ely cap u es empo al dependencies wi hin he ime se ies da a. The p edic ed oo con ac e en s se e as a ga eway o econs uc ing c i ical gai ea u es. Table 1: Summa ized lis o li e a u e showcasing he use o gai ea u es in Pa kinson’s disease p edic ion. Pape Yea Task Fea u e Iden i ica ion o mo o p og ession in Pa kinson’s disease using wea able senso s and machine lea ning [3] 2023 Classi ica ion o PD se e i y (MDS-UPDRS-III a ing scale) and iden i ica ion o impo an ea u es 122 gai ea u es Accele ome y-Based Digi al Gai Cha ac e is ics o Classi ica ion o Pa kinson’s Disease: Wha Coun s? [4] 2020 Classi ica ion o PD (non-PD, PD) and iden i ica ion o impo an ea u es 210 gai ea u es Gai and emo in es iga ion using machine lea ning echniques o he diagnosis o Pa kinson disease [5] 2018 Classi ica ion o PD (non-PD, PD) s ance ime, swing ime, and s ide ime Biome ic and mobile gai analysis o ea ly diagnosis and he apy moni o ing in Pa kinson's disease [6] 2011 Classi ica ion o PD (non-PD, ea ly PD, in e media e PD) 12 gai ea u es 2. Me hod 2.1 Da ase The Wea Gai -PD [7] da ase was used o his analysis. This da ase is one o he i s open- access da abase ha p o ides synch onized eco dings o aw IMU signals, insole p essu e senso da a, and p essu e-sensing walkway measu emen s o PD pa ien s. The da ase is composed o mul iple senso modali ies including IMUs placed a 13 body loca ions, insole IMUs, and walkway oo con ac in o ma ion ha p o ide bina y labels (1 o con ac and 0 o no con ac ) o le o igh oo con ac . IMU da a we e collec ed om he subjec s pe o ming eigh asks: sel -paced walking, hu ied pace, sel -paced walking on a ma , hu ied pace on a ma , sel -paced u ning on a ma , andem gai , he Timed-Up-and-Go (TUG) es , and balance. Fo he analysis, o he 123 subjec s, 20% (25 subjec s) was ese ed as an ex e nal es se . The da a om he emaining 98 subjec s we e used o aining, o which 14 we e se aside as an in e nal alida ion da a. 2.2 Model A chi ec u e Figu e 2: Model A chi ec u e whe e “WN” e e s o weigh no maliza ion, “Con ” e e s o con olu ional laye , and “ReLU” e e s o ec i ied linea uni ac i a ion. The 1D CNN-based neu al ne wo k model was used o p edic he ou pu (bina y p edic ion o le and igh oo con ac pe da apoin ) om he inpu ( ime se ies inpu o 6 o 12 channels). The inpu passes h ough se en laye s o 1-D CNN laye s wi h he classi ie laye a ached a he end o wo bina y p edic ions (le and igh oo con ac ). Skip connec ion was used o gi e he middle con olu ional laye s mo e con ex o he o iginal inpu [8-9]. De ailed model a chi ec u e is p esen ed in Figu e 2. 2.2 Pe o mance Measu emen Poin -wise accu acy o bina y classi ica ion was calcula ed o bo h pe o mance e alua ion measu emen , and aining loss. The accu acy was calcula ed using bina y c oss en opy whe e he loss L is de ined: 𝐿 =−1 𝑁&[ ! "#$ 𝑦"log(𝑝")+(1−𝑦")log(1−𝑝")0], whe e N is he o al numbe o samples, 𝒑𝒊 is he p edic ed p obabili y o con ac o he i h sample, and 𝒚𝒊 is he ue (g ound u h) label o he i h sample wi h i being he index o da a poin s. 3. Resul s 3.1 S ep Measu emen Accu acy Ou esul s demons a e ha ankle senso s yield highly accu a e p edic ions o oo con ac e en s, achie ing 97% poin -wise accu acy ac oss all es samples. Howe e , p edic ion pe o mance is less eliable when using senso s placed a o he body loca ions, highligh ing he impo ance o senso placemen in gai -based models. Table 2: Poin -wise accu acy pe senso inpu s (column) and p edic ion a ge ( ows) on he es se Ankle Lowe back W is Righ con ac 97.23 92.97 91.79 Le con ac 97.17 92.57 91.63 O e all 97.20 92.77 91.71 Figu e 3: Example plo s showing accele ome e [x, y, z] inpu s (blue, yellow, and g een espec i ely), le oo con ac p edic ion ( ed), le oo con ac g ound u h (black), and di e en ypes o mo emen deno ed as backg ound colo s. The plo s a e o ankle ( op), w is (middle), and lowe back (bo om) senso s. 4. Discussion Conclusion In his wo k, we p oposed a CNN-based s ep measu emen model which can be used o p oduce clinically meaning ul gai ea u es o PD p edic ions. Ou esul s demons a es ha he IMU senso s placed on ankle p oduce he bes esul when i comes o s ep measu emen wi h 97% poin -wise accu acy, ollowed by lowe back and w is placemen s. Ca ea s Despi e p omising p elimina y esul s, mo e in es iga ion is u he equi ed in model ine- uning and sensi i i y analysis. Model pe o mance can be u he imp o ed h ough se e al e inemen s wi h he ollowing app oaches. Fi s , smoo hing can be applied o e physiologically implausible “spikes” in p edic ed oo con ac when he p edic ed s ep du a ion is sho e han a p e-de ined h eshold. Second, sensi i i y analysis can be pe o med ega ding he p obabili y h eshold o decla ing a con ac e en (cu en ly se a 0.5). Finally, inco po a ing addi ional e e ence signals such as oo p essu e measu emen s could p o ide complemen a y in o ma ion o model imp o emen . Challenges The e emains challenge in adop ing he walkway con ac da a as he g ound u h in aining. Cases such as pa ial oo alls o edge cases, whe e a subjec ’s oo con ac ed he walkway pad only pa ially o he s ep is made en i ely ou o he pad, a e o en no coun ed as con ac in he walkway da a, whe eas he model may co ec ly de ec hem. This disc epancy complica es he e alua ion p ocess and necessi a es addi ional assump ions ega ding model pe o mance when compu ing accu acy me ics. A mo e lexible de ini ion o g ound u h may be equi ed o accoun o such cases. Fu u e Di ec ion Fu u e wo k in ol es alida ing he model on ex e nal da ase s ha lack explici con ac labels bu ha e highe -le el gai in o ma ion such as s ep coun s o du a ions. Gi en such da ase s, an e alua ion can be made be ween he o iginal gai ea u es and he ea u es ha a e de i ed based on ou s ep measu emen model. This ype o ea u e-le el alida ion would demons a e ou model’s gene aliza ion capabili ies ex ending o ex e nal da ase s and u he p o es ou model’s abili y as an e ec i e open-sou ce amewo k in c ea ing clinically meaning ul ea u es. Re e ences 1. Ni mal, K., and A. G. S eeji h. "Noise modeling and analysis o an IMU-based a i ude senso : Imp o emen o pe o mance by il e ing and senso usion." Ad ances in op ical and mechanical echnologies o elescopes and ins umen a ion II. Vol. 9912. SPIE, 2016. 2. Tedaldi, Da id, Albe o P e o, and Emanuele Menega i. "A obus and easy o implemen me hod o IMU calib a ion wi hou ex e nal equipmen s." 2014 IEEE in e na ional con e ence on obo ics and au oma ion (ICRA). IEEE, 2014. 3. So i akis, Cha alampos, e al. "Iden i ica ion o mo o p og ession in Pa kinson’s disease using wea able senso s and machine lea ning." npj Pa kinson's Disease 9.1 (2023): 142. 4. Rehman, Rana Zia U , e al. "Accele ome y-based digi al gai cha ac e is ics o classi ica ion o Pa kinson's disease: wha coun s?" IEEE Open Jou nal o Enginee ing in Medicine and Biology 1 (2020): 65-73. 5. Abdulhay, Enas, e al. "Gai and emo in es iga ion using machine lea ning echniques o he diagnosis o Pa kinson disease." Fu u e Gene a ion Compu e Sys ems 83 (2018): 366-373. 6. Ba h, Jens, e al. "Biome ic and mobile gai analysis o ea ly diagnosis and he apy moni o ing in Pa kinson's disease." 2011 annual in e na ional con e ence o he IEEE enginee ing in medicine and biology socie y. IEEE, 2011. 7. Ande son, An hony J., e al. "Wea Gai -PD: An Open-Access Wea ables Da ase o Gai in Pa kinson’s Disease and Age-Ma ched Con ols." medRxi (2024): 2024-09. 8. Zhou, Zongwei, e al. "Une ++: Redesigning skip connec ions o exploi mul iscale ea u es in image segmen a ion." IEEE ansac ions on medical imaging 39.6 (2019): 1856-1867. 9. D ozdzal, Michal, e al. "The impo ance o skip connec ions in biomedical image segmen a ion." In e na ional wo kshop on deep lea ning in medical image analysis. Cham: Sp inge In e na ional Publishing, 2016.