Early cerebral palsy risk screening through wearable accelerometry in infants under nine weeks

Abstract

The General Movement Assessment (GMA) according to Prechtl’s method has the best predictive value for early detection of Cerebral Palsy (CP) in infants under 5 months of age. However, access to specialists in this assessment is scarce, particularly in lower and middle-income countries. The main objective was to determine the better accelerometric features for distinguishing spontaneous movements of infants with risk factors for CP (RF), and healthy controls (HC), under 9 weeks. We carried out a cross-sectional study. General movements (GMs) were recorded in 48 infants under 9 weeks of age, 12 RF and 36 HC, using an instrumented assessment (accelerometers on limbs and trunk) and Prechtl’s method. Clinical variables and 62 accelerometer parameters were collected and analysed using descriptive and inferential statistics. To classify infants, we employed the Random Forest Classifier, based on their condition (Healthy vs. Risk Factors), with subsequent analysis of model accuracy. Afterward, we determined the features that best differentiated between the two groups. We found 46 parameters that differentiate RF and HC groups. Random Forest classified infants with 100% accuracy. Eight parameters were optimal for differentiation, and half of them were from trunk sensors. Wireless accelerometry effectively identified infant movement patterns indicative of cerebral palsy risk factors. This study establishes a scalable, personnel-independent screening of neonatal neurodevelopmental disorders, especially where specialized expertise is unavailable via Precht´s method. Deployment would facilitate widespread, cost-effective early risk stratification for CP.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-54620-y.

Introduction

Cerebral palsy (CP) is the principal cause of paediatric physical disability. Its diagnosis is complex and involves the presence of three elements: risk factors, motor dysfunction, and abnormal neuroimaging^1,2. There is no single test for CP diagnosis; therefore, definitive diagnosis necessitates evaluation by a professional highly specialized in the detection and interpretation of the clinical signs of cerebral palsy³. Considering this complexity, CP is often diagnosed during the first 2 years of life in high-income countries or later in middle- and low-income countries⁴. Early diagnosis allows early interventions that could improve the outcomes of children^2,5–9.

A widely validated diagnostic tool for early diagnosis of motor dysfunction is the General Movement Assessment (GMA), which is a standardized clinical evaluation and a reliable method, with the highest sensitivity and specificity in the diagnosis of CP in infants under 5 months of age. It has been recognized as the best clinical tool and internationally recommended to predict CP^7,10–13. Similarly, the assessment of risk factors for cerebral palsy is performed by trained clinicians with the same purpose, as a tool for early diagnosis. These risk factors are classified according to the time frame in which they occur and include multiple conditions present in mothers prior to pregnancy, complications during pregnancy and childbirth, as well as diseases in newborns and infants associated with reduced brain perfusion, which have been reported in various epidemiological studies^1,9,14–21. Despite the importance of these early diagnosis tools, several hospitals do not incorporate them into their screening programs.

GMA is an observational, qualitative evaluation based on the interpretation of a recorded short video of an infant moving spontaneously. These movements are called “general movements” (GMs). There are 3 periods of evaluation: preterm (until 40 post-menstrual weeks), writhing (between 40 post-menstrual weeks and 9 post-term weeks), and fidgety (from 6 to 20 post-term weeks). During the writhing period, spontaneous movement involve the whole body, varying in sequence, speed, and amplitude. GMs can be classified into four types: Normal Writhing, Poor Repertoire, Cramped Synchronized, or Chaotic²². The presence of CS movements has 97% specificity for future diagnosis of cerebral palsy^7,12, however outcomes related to other types of GMs are still unspecific²³. GMA is a technique based on the Gestalt perception of complexity and variability of movements and requires experience and certification of GM Trust®, a nonprofit organization that dictates international official courses²². Therefore, this assessment required highly trained professionals, who are scarce, particularly in low- and middle-income countries.

Recently, several groups have been developing automated techniques to perform quantitative analysis of GMs, using different sensors and artificial intelligence techniques²⁴; however, no standard method exists. Therefore, researchers are continuously testing new methodologies²⁵. Several studies have described the use of accelerometers to evaluate movements in infants, but the protocols are not comparable, differing in acquisition protocols, parameters extracted, and data processing techniques^26–39. Accelerometers measure three-dimensional motion, offering advantages such as small size, wireless capability, lightweight design, portability, low cost, and high temporal resolution^24,40,41.

Considering that in lower- and middle-income countries, GMA is not yet incorporated into high-risk follow-up programs, we designed a spontaneous movement assessment method using wireless accelerometers and computer-based analysis for infants under 9 weeks of age³¹, with the long-term goal of making this objective technology widely available and facilitating the early detection of sensorimotor impairments. The novelty of this study lies in the simultaneous use of six wearable accelerometers placed on both limbs and trunk, combined with classification according to Prechtl’s method and a Random Forest analysis, in infants under 9 weeks of age — an age range and sensor configuration not previously reported in the literature.

Our aim in the present study is to determine the accelerometric parameters that better distinguish spontaneous movements of infants with and without risk factors for cerebral palsy, under 9 weeks. Our hypothesis is that spontaneous movements of infants with risk factors for CP differ quantitatively from those of healthy controls in ways that can be objectively captured by multi-site wearable accelerometry, enabling discrimination between the two groups using accelerometric parameters.

Methods

Study design and population

We performed an analytical cross-sectional study. All infants were under 9 weeks post-term (age corrected as if born at 40 postmenstrual weeks). To obtain the population with risk factors for cerebral palsy, infants were recruited during their hospital stay at the “Luis Calvo Mackenna” Hospital, in the Neonatology Intensive Care Unit (NICU) and a basic ward, between November 2022 and March 2023. For the healthy population, we used data obtained previously from healthy infants born at “El Carmen” Hospital between August 2018 and November 2019.

Parents or legal guardians signed an informed consent form for all infants for both study participation and publication of deidentifying information. Informed consent was obtained from all subjects and/or their legal guardian(s). The study has been carried out in accordance with The Code of Ethics of the World Medical Association (Declaration of Helsinki).

Inclusion and exclusion criteria

The inclusion and exclusion criteria were based on the previous similar studies^27,36,39 and are detailed in Table 1 for both groups: Healthy Controls (HC) and those with Risk Factors for CP (RF).

Table 1.

Inclusion and exclusion criteria.

Consent and ethics

The study was approved by the scientific ethics committee of the Eastern area of Santiago, Chile (SSMO), and the Central area (SSMC #048975).

General movements recordings

Recordings consisted of a video of the spontaneous movements of infants using a GoPro 8 Hero Black camera, along with six “Wear Notch Pioneer kit” accelerometers. These were attached to the babies’ skin with hypoallergenic adhesive tape in the following locations: ankles (distal anterior leg, above malleolus), wrists (distal dorsal forearm, like a watch), pelvis (anterior midline, between Anterior Superior Iliac Spines), and thorax (at the center of the sternum). Caregivers were allowed to be present during the recording, but out of their infant’s field of view to avoid distracting them and interfering with their spontaneous mobility. One or more measurement attempts were made for each younger infant, until a recording of at least 2 min in duration that met the required conditions for correct recording according to GMA was obtained: awake, calm, without crying, without using a pacifier^11,42–44. Video recording was only used to classify the type of movement according to Prechtl’s method.

Data handling

We created a database with the population included in this study and consolidated it with the previous data from the healthy population³¹. We reviewed the data from the previous study and used only those that met the inclusion and exclusion criteria. The general movement assessment (GMA) of each infant in both groups was performed by professionals certified by the GM Trust® institution, in the advanced course, classifying them into the following categories: Normal Writhing, Poor Repertoire, Cramped Synchronized, and Chaotic²².

Clinical variables

We recorded the following information from the patients’ clinical records:

• Mother’s data: country of birth, age at the time of birth, region of residence, pregnancy characteristics.

• Birth data: sex, gestational age, weight, height, head circumference, Apgar score.

• Hospitalization data (only for the group with pathology): discharge diagnosis, detail of all risk factors below, relevant details of the physical examination, results of neuroimaging.

Risk Factors for CP^{7,15,16,19,45–48}. We assessed the following risk factors:

• Pre-conceptional: Maternal age (< 20 years, > 34 years), low maternal socioeconomic and/or educational level, first-time mother, short or prolonged intergestational interval, assisted reproductive technologies (such as in vitro fertilization), adverse obstetric history (miscarriage, stillbirth, neonatal death, premature birth), maternal systemic diseases (diabetes mellitus, thyroid disease, seizures, intellectual disability, coagulation disorders, immune system disorders), maternal drug use, stimulant use, and malnutrition, family history of cerebral palsy.

• Pre-natal: Multiple pregnancy, maternal systemic diseases during pregnancy (pregnancy-induced hypertension, preeclampsia, eclampsia), oligohydramnios, polyhydramnios, pregnancy complications: maternal and intrauterine infections (chorioamnionitis, TORCH infections), symptoms of preterm labor, premature rupture of membranes, placental flow disorders (placenta previa, placental abruption, placental insufficiency, intrauterine growth restriction, intrauterine hypoxia, vaginal bleeding).

• Peri-natal: Premature delivery (< 37 weeks) or post-term delivery (> 41 weeks), emergency cesarean section, complicated delivery: instrumental delivery (vacuum-assisted, forceps), cord prolapse, severe hemorrhage, placental abruption, constricted umbilical cord, severe malposition, intrapartum hypoxemic events (maternal shock or cardiorespiratory arrest, uterine rupture), perinatal asphyxia (hypoxic-ischemic encephalopathy or requirement for advanced resuscitation), low birth weight.

• Post-natal: Serious illnesses in the first week of life, stay in Neonatal Intensive Care Units, acute respiratory failure (Respiratory distress syndrome, pneumonia, bronchopulmonary dysplasia, congenital heart disease, meconium aspiration syndrome), artificial respiratory support (mechanical ventilation, oxygen therapy), severe and/or cerebral infections (such as sepsis, meningitis, encephalitis, necrotizing enterocolitis), major congenital malformations: cerebral (such as congenital hydrocephalus, microcephaly, microgyria, pachygyria, schizencephaly, porencephaly), cardiac (such as severe, cyanotic congenital heart disease), digestive (such as esophageal atresia, among others), respiratory (such as diaphragmatic hernia), neonatal brain injury (hypoxic-ischemic encephalopathy, intracranial hemorrhage, periventricular leukomalacia, Hydrocephalus, cerebral infarctions, traumatic injuries), Neonatal seizures, cardiovascular disorders (patent ductus arteriosus, hypotension, shock, cardiorespiratory arrest, need for vasoactive drugs, drowning), neonatal metabolic or endocrine disorders (hypoglycemia, hypothyroidism, severe hyperbilirubinemia, inborn errors of metabolism), ALTE (Apparent Life-Threatening Events), high-risk BRUE (Brief Resolved Unexplained Events), major neonatal surgery (cardiac and non-cardiac, such as esophageal atresia, gastroschisis, necrotizing enterocolitis, among others), use of corticosteroids before 7 days of life, Abnormal neonatal physical examination (in term newborns): abnormal fontanelle, abnormal tone, malformations congenital, altered consciousness (irritability or lethargy), impaired thermoregulation.

Data analysis

The features used and calculated for the analysis of baby movements were based on our previous study³¹, where the methodology and analytical pipeline were validated. Briefly, the accelerometer signal was processed using a bidirectional Butterworth 4th-order bandpass filter (0.1–20 Hz), and six categories of kinematic parameters were extracted for both trunk and extremities: acceleration and velocity, cross-correlation, skewness, kurtosis, periodicity, and area outside the mean of velocity. Outlier removal and normality testing were applied prior to statistical comparisons, as described below. Full methodological details are available in Marín-Palma et al.³¹.

For data processing, the accelerometer signal was first filtered using a bidirectional filter without offset (Butterworth 4th order, high pass of 0.1 Hz, and low pass of 20 Hz). Following the analyses developed by Meinecke et al.²³, Heinze et al.¹⁹, and Bultmann et al.¹⁷ we estimate 6 categories of variables (see Table 2) for trunk and extremities data (for details see³¹:

Table 2.

Parameters obtained to perform our analysis.

Parameters category	From extremities	From thorax and pelvis
Acceleration and velocity	Minimum acceleration of all extremities Maximum acceleration of all extremities Minimum velocity of all extremities Maximum velocity of all extremities Median acceleration of both upper extremities Median acceleration of both lower extremities Median velocity of both upper extremities Median velocity of both lower extremities	Minimum acceleration of the thorax Maximum acceleration of the thorax Minimum acceleration of the pelvis Maximum acceleration of the pelvis Minimum velocity of the thorax Maximum velocity of the thorax Minimum velocity of the pelvis Maximum velocity of the pelvis Median acceleration of the thorax Median acceleration of the pelvis Median velocity of the thorax Median velocity of the pelvis
Cross-correlation (CC)	Mean average of CC in acceleration between all extremities Mean average of CC of velocity between all extremities CC of acceleration between left and right lower extremities CC of velocity between left and right lower extremities CC of acceleration between left and right upper extremities CC of velocity between left and right upper extremities CC of acceleration between left upper extremity and left lower extremity CC of acceleration between left upper extremity and right lower extremity CC of acceleration between right upper extremity and right lower extremity CC of acceleration between right upper extremity and left lower extremity CC of velocity between left upper extremity and left lower extremity CC of velocity between left upper extremity and right lower extremity CC of velocity between right upper extremity and right lower extremity CC of velocity between right upper extremity and left lower extremity	CC of acceleration between thorax and pelvis CC of velocity between thorax and pelvis
Skewness	Skewness of the acceleration of the lower extremities Skewness of the velocity of the lower extremities Skewness of the acceleration of the upper extremities Skewness of the velocity of the upper extremities	Skewness of the acceleration of the thorax Skewness of the velocity of the thorax Skewness of the acceleration of the pelvis Skewness of the velocity of the pelvis
Kurtosis	Kurtosis of the acceleration of the lower extremities Kurtosis of the velocity of the lower extremities Kurtosis of the acceleration of the upper extremities Kurtosis of the velocity of the upper extremities	Kurtosis of the acceleration of the thorax Kurtosis of the velocity of the thorax Kurtosis of the acceleration of the pelvis Kurtosis of the velocity of the pelvis
Periodicity	Periodicity in the velocity of the lower extremities Periodicity in the velocity of the upper extremities	Periodicity in the velocity of the trunk
Area outside of standard deviation of moving average and area differing from moving average	Area in which the velocity profiles of the lower extremities are outside of the standard deviation of the moving average of the same lower extremities speed profile Area in which the velocity profiles of the upper extremities are outside of the standard deviation of the moving average of the same upper extremities speed profile Area in which the velocity profiles of the lower extremities differ from the mean average of velocity profile of the same lower extremities Area in which the velocity profiles of the upper extremities differ from the mean average of velocity profile of the same upper extremities	Area in which the velocity profiles of the trunk are outside of the standard deviation of the moving average of the same trunk speed profile Area in which the velocity profiles of the trunk differ from the mean average of velocity profile of the same trunk

1. Acceleration and velocity. Filtered 3D acceleration and the integral of the acceleration time series. For each segment, we combined the x, y, and z axes using the Euclidean norm of acceleration or velocity.

2. Cross-correlation. Measure of similarity between 2 signals, reflecting the degree of synchronization between 2 segments in time. Therefore, high values of cross-correlation should be interpreted together with other parameters.

3. Skewness of acceleration and velocity. Measure of the degree of symmetry of the distribution with respect to the mean. If the value is close to 0, the distribution is symmetrical; if it is positive, it means that the data accumulates to the left, and if it is negative, it means that the data accumulates to the right.

4. Kurtosis of acceleration and velocity. Measure of the data distribution’s tailedness in relation to the normal distribution. A positive Kurtosis indicates a greater concentration of similar values, while a negative Kurtosis indicates a distribution with wider and homogeneous values.

5. Periodicity of velocity. Quantified by the number of intersections of the velocity signal with the average of this signal calculated in windows of 1000 frames (moving window). The use of the periodicity was to determine the regularity and frequency of the motions in time.

6. Area outside the mean or standard deviation of velocity. Metric computed by subtracting the area of the velocity signal with the moving average or the moving standard deviation of the signal. We used a moving window of 100 frames. Its use was to estimate the amplitude variability of motions.

Table 2 contains all the estimated parameters for limbs and trunk.

Statistical methods

We used descriptive statistics to characterize the sample and inferential statistics to analyse accelerometer data, using Python 3.10.9 (Python Software Foundation, Beaverton, United States), with codes generated by our laboratory, to characterize the variables between groups.

Before the statistical analysis, we removed the outliers of each parameter using a threshold of 1.5 times the IQR above the upper quartile (75th percentile) and below the lower quartile (25th percentile). Then, the processed parameters were analysed using the Shapiro-Wilk test to check the normality of the data distribution. If data did not have a normal distribution, Mann-Whitney U tests were used to evaluate differences between groups, considering a p value ≤ 0.05, with Hedge as a post-hoc analysis. If the data had a normal distribution, we applied a T-test for independent samples.

Random forest analysis

To classify infants based on their clinical condition (Healthy Controls vs. Risk Factors for CP), we implemented a Random Forest classifier using the Scikit-learn library in Python. We divided the sample onto training (38 samples) and testing sets (10 samples). To ensure optimal generalization and prevent overfitting, we performed a systematic Grid Search. This process evaluated various combinations of hyperparameters, including the number of trees, maximum tree depth, and the minimum samples required to split a node. During the training phase, k-fold cross-validation was employed to stabilize the model’s performance estimates and refine the selection of the best-performing model. Subsequently, we calculated a confusion matrix and then evaluated the accuracy of our model.

Finally, to provide clinical interpretability, we performed a Feature Importance analysis, using used Recursive Feature Elimination with Cross-Validation (RFECV). This allowed us to rank the movement parameters based on their contribution to the classification accuracy. This step was crucial for identifying the variables that most effectively differentiate infants with risk factors for cerebral palsy from healthy controls.

Results

Clinical characteristics

We recruited a total of 36 healthy patients and 25 with Risk Factors for CP (valid recordings were obtained for only 12, as the remaining participants either could not be recorded during the required behavioral state (awake, calm, without crying, without pacifier), or did not meet the minimum recording duration of 2 min with adequate signal quality, required for valid GMA analysis^42,44. Of the 25 infants with risk factors who were recruited, 6 could not be recorded, and 7 additional recordings were excluded due to failure to meet these criteria.

Almost half of the total sample (47.9%) were female, and only 3 were born before 37 weeks of gestational age. The healthy control group assessments ranged from 0 to 5 weeks post-term, while the risk factor group ranged from 0 to 9 weeks.

The main diagnoses in the Risk Factors group were recovered neonatal depression, operated congenital heart disease, operated congenital diaphragmatic hernia, gastroschisis, necrotizing enterocolitis, acute respiratory failure, hypotonic syndrome, apneic syndrome, and tuberous sclerosis.

The details of the clinical characteristics of each patient are summarized in Supplementary material (1) Table 3 summarizes the number of risk factors in each patient, where each column represents a participant. As observed, infants in the Risk Factors group had between 6 and 11 risk factors, with a median of 9.5. The details of the risk factors for each patient are summarized in Supplementary material (2) Most frequent risk factors were postnatal: artificial respiratory support, acute respiratory failure, severe illnesses during the first week of life or stay in NICU, altered physical examination, major neonatal surgery, and major congenital malformations.

Table 3.

Distribution of risk factors in each infant of the RF group*.

Participants	#1	#2	#3	#4	#5	#6	#7	#8	#9	#10	#11	#12
N° of preconceptional risk factors	1	1	2	1	1	1	1	1	1	1	0	2
N° of Prenatal risk factors	0	2	0	1	3	0	0	2	0	1	1	1
N° of perinatal risk factors	1	1	1	0	2	0	1	2	1	1	2	2
N° postnatal risk factors	6	6	8	6	5	5	7	5	7	3	8	5
Total number of Risk Factors	8	10	11	8	11	6	9	10	9	6	11	10

*Each column represents an infant.

Median hospital stay for this group was 66 days, with an interquartile range of 33–82 days, and 11 patients required mechanical ventilation during their hospitalization, with a median of 4 days. Eleven babies had surgeries with general anaesthesia during their stay.

Neuroimaging was performed on 11 patients, of whom 5 had abnormalities, with 3 of these having affected the white matter. These results were homologated to the classification suggested by the Surveillance of Cerebral Palsy in Europe Association and are shown in supplementary material 3 (SM3).

Table 4 compares sociodemographic and clinical variables between the two groups. Variables were similar between both groups, except for birth weight, height, and head circumference, which is entirely expected, considering that the Risk Factors group includes premature infants.

Table 4.

Sociodemographic and clinical characteristics of the sample.

Variable	Healthy (N = 36)				With risk factors for cerebral palsy (N = 12)				Mann-Whitney- Wilcoxon test (p value)
	N (%)	Median (p25–p75)	Mean (Standard deviation)	Shapiro Wilk test (p value)	N (%)	Median (p25–p75)	Mean (Standard deviation)	Shapiro Wilk test (p value)
Mother’s history
Born in Chile	33 (91.6%)				10 (83.3%)
Age at the baby’s birth		25.5 (23–30)	26.2 (5.4)	0.903		33 (27–37)	32.1 (6.6)	0.049	0.787
High-risk pregnancy	0 (0%)				11 (91.6%)
Birth’s data
Female	18 (50%)				5 (41.7%)
Gestational age (weeks)		39 (38–39.3)	39 (0.81)	0.000		37.5 (36.8–38)	36.8 (2)	0.011	0.618
Birth weight (gr)		3440 (3120–3714)	3437 (445)	0.712		2869 (2560–3276)	2949 (612)	0.999	0.007
Birth size (cm)		50 (49.5–51)	49.9 (2.3)	0.065		47 (45.5–48.8)	47.4 (2.4)	0.070	0.002
Birth cranial circumference (cm)		34.5 (34–35)	34.7 (1.1)	0.170		33.3 (32.6–34.6)	33.4 (1.4)	0.639	0.024
Apgar 1 min		9 (8.8–9)	8.72 (0.51)	0.015		7 (4.8–8)	6.2 (2.7)	0.000	0.471
Apgar 5 min		9 (9–9)	9.2 (0.42)	0.754		8.5 (7–9)	8 (1.5)	1000	0.610

Analysis of general movements according to the Prechtl’s method

Patient movements for the total sample were classified into only 2 categories: Normal Writhing and Poor Repertoire. No cramped synchronized or chaotic movements were observed.

In both groups, most participants exhibited movements classified as PR. In the Healthy group 25 infants (69%), while in the Risk Factor group, there were 10 (83%).

Differences in accelerometer parameters between both groups

We obtained the 62 accelerometer parameters for each patient as detailed in Table 2. 46 parameters were statistically different between the groups (Fig. 1 and Table SM4).

Fig. 1.

Comparison of trunk and limb parameters by group. (A) Trunk parameters. (B) Limb parameters. Significant differences between groups are depicted in bold font (p values in supplementary material 4). T=Thorax, P=Pelvis, UE= Upper Extremities, LE=Lower extremities, R=Right, L=Left, B=Both(R & L), vel=velocity, acc= acceleration. For comparison purposes, the values are centered (mean equal to zero) and scaled (1 standard deviation equal to one).

Classification using random forest algorithm

The Random Forest classifier demonstrated high discriminative power in distinguishing between infants in the Healthy Control (HC) and Risk Factor (RF) groups. After hyperparameter optimization via grid search, the model achieved an overall accuracy, sensitivity and specificity of 100%, namely, therefore our dataset can be classified perfectly in two groups (Fig. 2A).

Fig. 2.

(A) Accuracy of the Random Forest model. The Confusion Matrix shows that classification was exact for HC (Label 0, N = 36) and RF (Label 1, N = 12). (B) Selected Features and Importance to distinguish between the two groups.

Consequently, 8 features had the best accuracy in the Random Forest Model. From higher to lower importance these variables were, Maximum acceleration of all extremities, Minimum acceleration of the thorax, Periodicity in the velocity of both upper extremities, Minimum velocity of the thorax, Minimum velocity of all extremities, Median acceleration of the pelvis, Kurtosis of the acceleration of both lower extremities and the Skewness of the velocity of the pelvis. Figure 2B shows the importance scores of the selected features.

Discussion

Our study is novel in comparing groups with and without CP risk factors in infants using accelerometers. We found 46 parameters that differentiate RF and HC groups, with 100% accuracy using Random Forest. We identified 8 parameters that best explain the difference between healthy infants and those at risk of cerebral palsy, with 4 of these parameters related to the trunk. The importance of the trunk is consistent with our previous publication³¹. In our current study, no participants exhibit neither cramped nor chaotic movement. It is essential to note that these abnormal movements have distinctive characteristics that are usually distinguishable in clinical practice, so accelerometry would likely not provide a significant contribution to screening. Therefore, the value of our methodology would potentially provide a tool for screening infants with mild sensorimotor impairments.

We decided to analyse movements in infants under 9 weeks because it is a stage where movement patterns are quite automatic and similar²². Physiologically, they undergo short periods of wakefulness with less mobility than at older ages, making the early detection of alterations more difficult, resulting in delayed diagnosis. Additionally, we chose accelerometers because they were portable and low-cost, having materials that allow for proper disinfection, making it a feasible technique to implement in an intrahospital context. Regarding the recording time, we opted for short-duration videos, emulating the requirements of Prechtl’s method^22,31.

Among the comparative advantages of our analysis is the use of six accelerometers. Including the upper and lower trunk allowed us to assess the contribution of this body segment to spontaneous movements, which is a remarkable advantage over other studies that only analyses limbs or include sensors on the head^{27,30,32,35,36,39}. However, we did not find descriptions in the literature regarding trunk parameters in infants at risk of cerebral palsy.

Comparing our selected parameters with other studies, we note that most of them are related to speed and acceleration. Heinze et al.³⁶ described additional statistical parameters (skewness and cross-correlation), which have a more challenging interpretation in clinical practice. In contrast to the findings of Meinecke et al.³⁹, feet parameters were not present in our selected parameters. Additionally, these studies do not include the classification of infants according to the type of general movement. As described by Meinecke et al.³⁹, when defining and calculating these parameters, areas outside the average and cross-correlation are associated with the presence of Cramped Synchronized movements.

Regarding the accuracy of our model, it is noteworthy that our accuracy is comparable to that obtained previously^36,39.

Our findings highlight the feasibility of extrapolating these data to clinical evaluation, recommending the incorporation of wearable technology and closer observation of hands and trunk movements when conducting spontaneous movement assessments.

Regarding the RF group, we were surprised that 50% had at least 9 risk factors, and most were postnatal factors associated with clinically severe conditions such as NICU stay, respiratory insufficiency, and use of mechanical ventilation. This alerts us to the need for more frequent and precise screening of babies hospitalized in our health centers. Amongst our study population, most patients were not clinically diagnosed as high risk for CP. Thus, it highlights the importance of follow-up and reporting our results as these risk factors were not apparently prioritized when screening for neurological abnormalities.

Considering all the above, the use of this technique could significantly contribute to the early identification of at-risk populations, particularly in contexts where trained personnel are unavailable to perform clinical evaluations according to Prechtl’s method.

This is an area of research with high potential, and the challenge now would be to delve deeper, evaluate consistency by increasing the sample size, and conduct medium-term follow-up of infants. It is imperative to continue conducting research in this area to advance toward the development of accessible, mass-use accelerometer-based devices for early screening. Such advances would optimize diagnosis and timely intervention, ultimately improving the functional prognosis of affected children.

Limitations

While these results are very promising given their feasibility and clinical applicability, we must be cautious when extrapolating and interpreting these data, as the sample is small and imbalanced from a national high-complexity center. Additionally, patients were recruited in the early stages of life, and our cross-sectional design involved only one measurement per patient, without follow-up, which does not allow for a definitive diagnosis of neurological pathologies. Finally, although most of our patients underwent neuroimaging, the majority were reported as having no pathological findings. This could be because most were brain ultrasounds performed early after insults, and there may not have been enough time for a structural lesion to develop.

Furthermore, our sample presented only two GM categories (Normal Writhing and Poor Repertoire), which limited the ability to explore the relationship between all four Prechtl GM types and the accelerometric parameters; future studies with larger and more diverse samples should address this gap.

Additionally, the absence of an independent external validation cohort is an important limitation; the 100% accuracy obtained should be interpreted as proof-of-concept, and replication in larger independent cohorts and with follow-up cohorts is required before clinical translation It should also be noted that the possibility that some Healthy Control infants may later develop CP cannot be excluded, given that nearly 50% of children with CP are born without overt risk factors⁴³.

Conclusions

Wireless accelerometry can distinguish between spontaneous movements of healthy and at-risk infants, aiding in the early detection of patients with risk factors for cerebral palsy.

Accelerometry holds promise for early diagnosis of neurodevelopmental disorders, supporting need for further research and integration into clinical practice. Their potential benefit includes improved diagnostic accuracy and intervention timeliness, ultimately improving the functional prognosis of affected infants.

Electronic Supplementary Material

Below is the link to the electronic supplementary material.

Author contributions

J.A.: conceptualization, methodology, data interpretation, visualization, writing—original draft.F.A.: Data collection, writing—review and editing.G.M.: Data collection, writing—review and editing.M.M-P.: Conceptualization, data interpretation, methodology, writing—review and editing.I.R-S.: Conceptualization, data interpretation, methodology, writing—review and editing.J.B-C.: Conceptualization, data interpretation, methodology, writing—review and editing.R.Z-I.: data interpretation, resources, writing—review and editing.R.J.: data interpretation, resources, writing—review and editing.P.I.B.: Conceptualization, methodology, supervision, data analysis, data interpretation, resources, writing—original draft.

Funding

The study was partially supported by Contec #78959 Corfo and FONDEF Grant # ID23I10437 from the National Agency for Research and Development (ANID), Government of Chile.

Data availability

The data reported here could be available at reasonable request and within relevant legal constraints. Please contact Pablo Burgos (pburgos@uchile.cl) for requests.

Declarations

Competing interests

The authors declare no competing interests.