Many neurological conditions affect the sensorimotor control of speech movements (eg, PD) or cognitive processes, specifically memory and language and perceptual processing (eg, AD and MCI). While thought formulation and motor planning are distinct in speech production [114], alterations in motor planning can occur alongside cognitive impairments, as discussed in various stages of AD [115]. This can cause acoustic changes in the physical speech signal.

Among the neurological diseases, PD (40/72, 56%) was the most investigated disorder, followed by AD and cognitive impairment (12/72, 17%). Other investigated disorders were MS (5/72, 7%), ALS (4/72, 6%), mild traumatic brain injury (mTBI; 3/72, 4%), Huntington disease (HD; 2/72, 3%), and ASD (2/72, 3%). A total of 4 articles focused on clinical purposes related to apathy (n=1, 25%); intellectual disability (ID; n=1, 25%); essential tremor (ET; n=1, 25%); and a group of central nervous system disorders (CNSDs), including HD and PD (n=1, 25%).

Disease diagnosis, differential diagnosis, severity assessment, and treatment monitoring were the most mentioned clinical purposes in the studies. Figure 4 depicts the distribution of articles with their clinical purposes according to disease categories. Some studies (18/72, 25%) addressed multiple clinical purposes, such as diagnosis and differential diagnosis or diagnosis and severity assessment. Disease diagnosis was widely researched. These studies focused on prevailing clinical challenges such as the lack of definitive objective biomarkers, the need for noninvasive biomarkers, challenges in discriminating diseases with similar symptoms (differential diagnosis), and challenges faced by vulnerable populations such as older adults or rural populations.

PD was the most investigated neurological condition in the studies (40/72, 56%). PD diagnosis, differential diagnosis, treatment monitoring, and severity assessment were explored. Many studies (28/40, 70%) investigated the diagnosis of PD by discriminating parkinsonian speech compared to healthy speech [42-69]. Several studies (3/40, 8%) carried out statistical comparisons of speech features in healthy and PD groups [70-72]. PD severity scores were predicted in the study by Viswanathan and Arjunan [73] for patients with PD and healthy controls. Early detection of PD was investigated in some studies (4/40, 10%) [42,47,57,65], in which participants with early PD were recruited to compare them with healthy participants. Early and mid-aged patients with PD were recruited in the studies by Montaña et al [65] and Wang et al [42], whereas the study by Lim et al [47] focused on patients with early- and advanced-stage PD according to the Hoehn and Yahr staging scale. Patients with PD were assessed in the study by Jeancolas et al [57] if they had been diagnosed with PD within 4 years before the study.

Some studies (4/40, 10%) investigated discrimination of similar-symptom diseases for differential diagnosis. For example, the study by Song et al [44] investigated distinguishing ataxic and hypokinetic dysarthria, which are commonly prevalent in neurodegenerative diseases. Patients with PD and cerebellar ataxia were considered representative cases of diagnosis. A statistical comparison of PD speech and MS speech was performed in the study by Vizza et al [72]. The challenge of differential diagnosis was addressed in the studies by Das et al [74] and Li et al [75] for atypical parkinsonian syndromes (APS) of progressive supranuclear palsy and multiple system atrophy.

In addition to diagnosis, the studies investigated PD severity predictions using the Unified Parkinson’s Disease Rating Scale (UPDRS) and Hoehn and Yahr scale. The studies by Viswanathan and Arjunan [73] and Hemmerling and Wojcik-Pedziwiatr [76] predicted UPDRS scores from the speech of patients with PD considering their medication status. Furthermore, Zhang et al [61] and Sztaho et al [67] predicted PD severity level from speech features in relation to the UPDRS and Hoehn and Yahr scale, respectively. The severity of PD symptoms was assessed in the study by Tunc et al [77] through collected speech samples at different points after taking levodopa.

The studies analyzed the speech of patients with PD to assess the impact of antiparkinsonian treatments on speech and voice function. For example, Suppa et al [43] studied speech changes during highly effective and less effective periods from levodopa therapy for PD. PD severity using UPDRS scores was predicted in the study by Hemmerling and Wojcik-Pedziwiatr [76] at different time points after taking levodopa. The effects of dopaminergic medication on speech function were investigated in the studies by Vandana et al [78] and Jain et al [79]. The impact of assistive speech devices in treating speech impairment was studied by Gaballah et al [80,81] considering these devices’ treatment capability outside the clinical facility. In these studies, Gaballah et al [81] investigated discrimination of PD speech and healthy speech under different environments and amplification conditions, whereas Gaballah et al [80] predicted the perceived voice quality of patients with PD with and without assistive speech amplifier devices.

Cognitive decline was assessed through speech analysis, including clinical conditions such as AD, dementia, and MCI. Dementia presents different symptoms at different severity levels of cognitive decline. MCI represents an early stage of cognitive decline without interference with everyday life but can act as a transition stage between healthy aging and dementia when it acts as preclinical AD [116]. AD can evolve over a continuum from normal cognition to MCI due to AD, followed by more severe AD dementia [117]. AD is identified as the most common form of dementia [117].

The studies in this review analyzed speech at different cognitive decline stages for diagnosis, differential diagnosis, and severity assessment (12/72, 17%). The focus was on discriminating impaired cognition from normal cognition. The discrimination of patients with AD [82] and MCI [83-85] from healthy controls was explored. Nagumo et al [85] studied differential diagnosis with global cognitive impairment as they considered patients with MCI, global cognitive impairment, and both forms of cognitive decline. On the other hand, some studies (3/12, 25%) [86-88] investigated both diagnosis and differential diagnosis of AD compared to patients with MCI and healthy participants. Automatic differentiation of patients with MCI and early dementia from healthy controls was studied by Bertini et al [89].

Differential diagnosis of AD and dementia with Lewy bodies was investigated in the study by Yamada et al [90], whereas Sumali et al [91] worked in discriminating between patients with depression and dementia as certain mental disorders (eg, depression) can cause pseudodementia, a temporary decline in mental cognition. While Al-Hameed et al [92] discriminated speech from neurodegenerative diseases, including patients with AD, MCI, and dementia, from speech from functional memory disorder, König et al [93] predicted neuropsychiatric inventory scores in a sample of patients with MCI through speech.

MS is a chronic inflammatory disease of the central nervous system that affects cognitive and motor functions causing motor and sensory impairments, visual disabilities, cognitive disorders, and speech and language deficits [118]. Different studies (5/72, 7%) investigated MS diagnosis and severity assessment through speech analysis. Discriminating patients with MS from healthy controls was investigated in several studies (4/5, 80%) [94-97], whereas Fazeli et al [94] explored the relationship between selected speech quality indexes and MS severity. Speech analysis among patients with MS at 2 different disease stages, in addition to comparisons of patients with MS versus healthy patients, was conducted in the study by Vizza et al [98].

ALS is a progressive motor neuron disease that affects upper and lower motor neurons in the motor cortex, the brain stem, and the spinal cord. ALS leads to muscular weakness and spasticity that can result in difficulties with mobility, breathing, and motor speech production [119]. Research on speech in ALS concentrated on identifying differential diagnoses and detecting bulbar involvement (4/72, 6%). Differentiation of speech from 3 participant groups, including patients with ALS with and without bulbar involvement and healthy controls, was conducted in the study by Tena et al [99]. The studies by Illa et al [100] and Likhachov et al [101] investigated discrimination of speech from patients with ALS and healthy participants. The ALS populations in both studies had shown signs of bulbar involvement. Furthermore, Mallela et al [102] investigated the diagnosis of ALS and differential diagnosis of ALS from PD by including ALS patients with a range of speech dysfunction severities.

Compared to other progressive neurological diseases that occur due to a physical or mental phenomenon originating within the body, mTBIs or concussions are initiated when a person experiences an external force to the head, causing some alteration in brain function [120]. Therefore, the studies in this review on concussion (3/72, 4%) were able to access baseline speech recordings from highly vulnerable populations such as athletes along with the postinjury speech and then aimed to discriminate concussed speech from that of healthy controls [103] and from the individuals’ healthy baseline speech [104]. Concussion detection through baseline, postconcussion, and posthealthy speech comparison was also studied by Wall et al [105].

HD is a rare severe neurodegenerative disease with a known natural history due to inheritance [106]. One study assessed the severity of HD, whereas the other study investigated the expression of emotions through vocal characteristics in participants with HD and pre-HD using emotion elicitation via speech. Riad et al [106] investigated the prediction of HD severity scores from speech features of a sample of participants with pre-HD and HD. Furthermore, they studied the relationship between speech variables and striatal volumes as well. In addition, the emotion expression capability of patients with HD was studied through speech-based emotion recognition [107].

ASD is a neurodevelopmental disorder that causes difficulties in social communication, particularly in verbal and nonverbal communication [121]. Most people living with ASD exhibit speech and expressive language abnormalities at different levels [121]. The 3% (2/72) of ASD articles in this review focused on discriminating ASD speech from healthy speech [108,109] and estimating autism severity by predicting Autism Diagnostic Observation Schedule scores from speech features [109]. It was noted that comorbid conditions were present in these populations, including attention-deficit/hyperactivity disorder [108] and children with suspicion of ASD, such as children with other language or developmental delays [109], in addition to children with typical development.

The neurological conditions that were the focus of other studies included ET (1/72, 1%), apathy (1/72, 1%), ID (1/72, 1%), and a broader category of CNSDs (1/72, 1%).

ET is among the most common tremor syndromes, which can encompass voice tremor as well [122]. To complement conventional neurological examination–based assessments, voice tremor in patients with ET was studied by Suppa et al [110] by discriminating the speech of patients with ET who did and did not manifest clinically overt voice tremor. Moreover, they discriminated between patients at baseline and after having medical treatments [110]. Apathy is identified as a motivation disorder that can present in several psychiatric and neurological conditions. Discrimination between patients with and without apathy was done in the study by König et al [111] through speech from patients with mild to moderate neurocognitive disorders. ID is a neurodevelopmental disease similar to ASD that causes cognitive delays in early childhood, resulting in delays in adaptive function, language, and speech [123]. Speech samples from children with typical development and those with ID were compared in the study by Aggarwal et al [112]. Lauraitis et al [113] investigated discrimination of speech impairment in patients with early-stage CNSDs from healthy speech. They included patients with HD, PD, cerebral palsy, stroke, and early dementia in their CNSD population.

Across the 18.5% (72/389) of neurological studies, speech data were captured through a range of speech tasks, from simple production of sounds or words to spontaneous speech in conversations. The characteristics of the speech tasks impact the reliability of speech features [124]. For example, speech quality features from continuous speech may be less reliable compared to more structured sustained phonation tasks [124]. Therefore, speech tasks and speech features together define speech data in clinical speech analysis [14].

Speech tasks across the 18.5% (72/389) of neurological studies can be categorized along a continuum from highly constrained to naturalistic speech elicitation. At the most structured end are sustained phonations and diadochokinetic tasks. In the middle range are reading and picture description tasks, whereas prompted speech tasks elicit the most naturalistic speech production.

Highly constrained tasks frequently appeared in studies of neurological diseases. These included sustained phonation of vowels (/a/, /e/, /i/, /o/, and /u/) and sounds such as “ah,” “eh,” “iuh,” and “iamh.” Similarly structured were oral diadochokinetic tasks, including alternating motion rate and sequential motion rate tasks. Alternating motion rate tasks require rapid repetition of a single syllable (eg, /pa/, /ta/, or /ka/), whereas sequential motion rate tasks involve sequences of different syllables (eg, /pa-ta-ka/).

Semistructured tasks provided a balance between control and naturalistic speech production. Reading tasks typically included reading a set of words, sentences, and passages. Picture description tasks elicited spontaneous speech within a potentially limited vocabulary set associated with the provided visual aid. Activity-related speech included speech fluency tasks, recall and summary tasks, and number-related tasks such as counting and subtraction. On the other hand, prompted speech tasks encourage more free and spontaneous speech within a guided procedure. Monologues on own experiences, conversations including neurological examinations, and spontaneous speech induced by interactive questions were some examples. It was also common for the studies to use multiple speech tasks, combining structured and semistructured tasks. The studies used speech tasks from culturally adapted test batteries such as Thammasat-National Electronics and Computer Technology Center (NECTEC)-Chula’s Thai Language and Cognition Assessment [86].

Overall, structured speech tasks were more widely used for neurodegenerative diseases such as PD and ALS, whereas semistructured speech tasks were common for cognitive impairment–related disorders such as AD, MCI, and dementia. Textbox 2 shows the application of different types of speech tasks in the studies with examples.

The studies extracted speech features based on the characteristics of the speech tasks. From a signal processing perspective, the speech features included in the studies can be broadly categorized into 3 main types: fundamental and advanced signal processing–based speech features and audio images (Textbox 3). The primary difference between fundamental and advanced signal processing–based speech features lies in their capacity to correspond to phonetic aspects of speech production. This distinction is important for understanding how these speech features are explained from both a speech science and clinical perspective. Being single-dimensional, both these types of speech features are suitable candidates for statistical analyses, traditional ML approaches, and deep neural networks (DNNs). On the other hand, audio images, a form of multidimensional speech representation, act as candidates for image-specialized neural networks such as convolutional neural networks (CNNs).

Speech features can be analyzed through 2 complementary lenses: their representation dimension during speech feature extraction and their physiological phonetic characteristics, which are derived from the field of speech science. Speech signals can be represented and features can be extracted in the time, frequency, time-frequency, and cepstral domains. From the physiological phonetic viewpoint, speech features capture biological and anatomical aspects of speech production, including articulation, phonation, prosody, and speech quality. Within the studies, speech articulation was assessed through time-domain features (eg, AMR [75], diadochokinetic rate, and diadochokinetic period [103]), frequency-domain features (eg, formants [42,76], Bark band energy-based features [42,54], spectral moment, and power spectral moment [76]) as well as cepstral domain features (eg, MFCC [42,45,54]). Similarly, speech phonation was assessed in features from time domain (eg, jitter, shimmer [42,45,62,73,76], amplitude and speed-based glottic cycle features [62,73], Teager-Kaiser energy operator based features [45,77], energy [76]), frequency domain (eg, fundamental frequency [42,48,62,73,76], bark band energy-based features [48], harmonic-to-noise ratio [45,73], noise-to-harmonic ratio [45]) and cepstral domain (eg, MFCC [48,53], linear-frequency cepstral coefficients [53], gamma-tone cepstral coefficients [53], cepstral peak performance [73]). Pause-based features [109] and fundamental frequency [109] are some examples for time domain and frequency domain features for prosody and rhythm assessment.

Different structures have been proposed for categorizing these features. Speech features were categorized as voicing, articulation, and prosodic features in the study by Li et al [75], whereas Wang et al [42] grouped speech features as phonatory, articulatory, prosodic, and cognitive-linguistic features. Phonatory features modeled abnormal patterns in the vocal fold vibrations, whereas articulation features captured deficits in articulatory movements of the lips, tongue, and jaw. Prosodic features, such as speech rate and timing, investigated paralinguistic aspects such as emotions, whereas cognitive-linguistic features considered vocabulary, phrase construction, and word repetitions [42]. The domain of speech feature representation pertains to the complexity and resource demands for feature extraction, whereas phonetic elements relate to speech task characteristics and the biological process involved in speech production.

Speech features are also categorized as linear or base speech features (eg, fundamental frequency, jitter, and shimmer) and nonlinear features, which are mostly derived from advanced signal processing techniques. The studies by Zhang et al [61] and Tunc et al [77] investigated nonlinear speech features such as correlation dimension, recurrence period density entropy, detrended fluctuation analysis, and pitch period entropy. Tunable Q-factor wavelet transform and empirical mode decomposition–based speech features are some more advanced speech features [77]. Different spectrograms, including mel spectrograms [44,54], linear spectrograms [54], and constant-Q transform spectrograms [54] are examples of speech image representations assessed in the included studies.

Secondary voice indexes were another type of speech feature available in the studies. They were indexes defined from other primary speech features. Example measures included the Dysphonia Severity Index, formant centralization ratio, and vowel metrics. The Dysphonia Severity Index is calculated from phonation time, jitter, fundamental frequency, and intensity, whereas the latter 2 are calculated from the measures of formants of vowels [72,94]. Some studies used standard speech feature sets, for example, the extended Geneva Minimalistic Acoustic Parameter Set [107], INTERSPEECH2016 Computational Paralinguistics Challenge speech feature set [49,110], and the emobase feature set [86]. Speech feature embeddings derived from a deep learning (DL) approach were also a type of speech feature representation assessed in the included studies [50,57].

Descriptive and predictive studies were included among the 18.5% (72/389) of neurological studies. Descriptive studies assessed the relationship between clinical variables and speech features (eg, whether speech features are associated with a disease or its severity), whereas predictive studies investigated the estimation of clinical variables from speech features (eg, predicting disease severity from speech features). Statistical analysis was largely applied in descriptive studies, whereas predictions were made through traditional ML and DL approaches.

Univariate, bivariate, and multivariate analyses of speech features were observed across the 18.5% (72/389) of neurological studies. Univariate analyses investigated descriptive measures and the statistical significance of speech features in differentiating pathological and healthy groups. Descriptive statistics of speech features within first hour and after 12 hours of medication of PD [78] and at 2 stages of MS [98] are some examples.

Further statistical significance tests showed the differences between speech features, mainly between healthy and impaired speech due to neurological health conditions. Several statistical significance tests were applied depending on the relationship between discriminating groups and the normality of the variables. For example, Suppa et al [110] statistically compared speech with and without voice tremors using the Student t test, whereas the paired Student t test was used to compare speech features before and after medication. To differentiate AD, dementia with Lewy bodies, and healthy speech, Yamada et al [90] compared speech features between groups after controlling for medication using 1-way analyses of covariance. Both studies explored the relationship between speech features and disease characteristics before building predictive models. The studies used bivariate analyses to explore the association between speech features and clinical variables, such as the correlation between speech measures and UPDRS scores for PD severity [73,76,77]. The study by König et al [111] also assessed the correlation between speech properties and the Apathy Inventory subscales to assess the properties of apathetic speech in people with cognitive disorders. Multivariate analysis was applied to simultaneously explore the relationship among clinical variables, speech features, and other confounding factors such as age and sex. For example, Svoboda et al [96] applied multiple linear regression of speech features, age, and sex to differentiate MS from healthy speech.

Table 1 summarizes the statistical analysis approaches applied in the selected studies.

The studies used traditional ML, DL, and hybrid approaches to predict clinical variables from speech features. Disease diagnosis was formulated as a binary classification, whereas differential diagnosis was extended into multi-class classifications. The severity assessment was primarily treated as a regression problem, whereas treatment monitoring involved either regression or classification of patients based on medication status. Classification problems were the most commonly addressed.

Among traditional ML classifications, the following were the most common algorithms: support vector machine, k-nearest neighbor, random forest, decision tree, Extreme Gradient Boosting, multilayer perceptron, and logistic regression. Linear regression, support vector regression, random forest regression, and artificial neural network regression were among the traditional ML regression techniques, for example, in Autism Diagnostic Observation Schedule score prediction for autism severity assessment [109] and prediction of neuropsychiatric inventory scores [93].

DL approaches were applied as deep feature extractors and end-to-end predictors. In deep feature extractors, DL capability was explored in self-extracting efficient feature representations from speech features in a supervised or unsupervised manner. The studies retrained special DL models and built their own models as feature extractors. Example deep speech feature extractors included the implementation of a standard DNN in the study by Gosztolya et al [97] and an autoencoder network in the study by Bertini et al [89]. Regarding the development of end-to-end DL models, recurrent neural network models and CNNs were among the most promising ones. Different architectures such as CNN with a modified Hybrid Mask U-Net architecture with an adaptive custom loss function for PD assessment [46] and bidirectional long short-term memory neural networks [105] were implemented. Hybrid architectures such as CNN–long short-term memory networks [102] and personalized convolutional recurrent neural networks [79] were also assessed in the studies. Transfer learning was applied to train computer vision–based approaches for clinical speech analysis. Examples included CNN14 [44] and AlexNet-based CNN for PD assessment [53].

There were fewer research attempts to explore DL approaches for clinical score prediction. Autism severity score prediction using DNN and CNN [109] and PD severity prediction using DNN [67] were among the few examples. Table 2 summarizes the ML and DL approaches applied in the selected studies.

Some studies (24/72, 33%) integrated multiple data science approaches. They first applied statistical techniques to screen and filter speech features before feeding them into ML-based prediction models. Application of unsupervised learning techniques was the least common, with only one study exploring unsupervised clustering. The study applied statistical analysis, unsupervised learning, and supervised classification to address the classification of patients with depression and dementia based on speech, demonstrating comprehensive use of data science approaches in the speech analysis pipeline [91].

We adapted the research process proposed in the study by Offermann et al [134] to operationalize research in design science. The proposed research process is structured in 3 main phases—problem identification, solution design, and evaluation—supporting both quantitative and qualitative research methods [134]. To build our research framework, we identified key activities and desired outcomes at each phase, focusing on research using primary speech data collection. Table 3 shows the research process, including proposed activities under each subprocess of the main stages of problem identification, solution design, and evaluation. Expected outcomes are proposed at each stage to guide the research progress.

Problem identification defines the diseases of interest, clinical purpose, and potential clinical applications. Furthermore, it helps characterize the speech impairment associated with the disease to formulate research hypotheses.

As per this review’s findings, digital clinical speech analysis addresses different clinical problems. For example, speech features are being researched as objective biomarkers for disorders such as AD, mTBI, and PD in cases in which a definite biomarker is not available. Moreover, speech analysis aims to empower the remote monitoring of patients with PD who are mostly older adults. Speech analysis tries to differentiate overlapping symptoms in different diseases, such as cognitive impairments and symptoms of natural aging. Therefore, researchers can focus on a particular health condition considering existing clinical challenges and evidence of associated speech impairments [114]. To explore clinical challenges and potential speech changes from diseases, existing literature and experiences from health care experts such as clinicians, speech-language pathologists, and frontline health professionals can be referred. Experiences from patients might also be beneficial when the target end applications are considered. Findings of this phase identify a research gap and define research objectives and research questions. Furthermore, the researcher can qualitatively characterize speech impairments and define research hypotheses for empirical research.

In the solution design phase, specific literature research on current clinical practices and associated clinical speech assessments is beneficial to identify the ground truth [114] and potential speech characteristics to focus on. Study design approaches and state-of-the-art data analysis techniques can be used to ensure the scientific validity and novelty of the research. Clinical problems can be mapped into a predictive modeling problem, and an appropriate speech data collection plan can be developed. Data collection should include patients’ demographics, assessment of health conditions through standard clinical measures, and speech recordings. Furthermore, it should define speech recording instances, conditions, recording equipment, and speech tasks. Recording instances define when to collect speech from the participants, such as before or after medication. Recording can be done in either controlled or uncontrolled environments based on the research objectives. For example, if the research objective is the remote monitoring of patients, recordings in an uncontrolled environment would be more appropriate. Speech tasks should be carefully selected to capture appropriate speech characteristics associated with the health condition.

We recommend a comprehensive analysis of speech features to conduct association analysis to clinical parameter predictions. The data analysis workflow can be divided into three main stages—primary analysis, secondary analysis, and tertiary analysis—to (1) explore participant characteristics and extract suitable speech features from speech signals, (2) examine the relationship between speech features and clinical variables, and (3) develop and optimize ML models to predict clinical variables from speech features.

The primary analysis quantifies speech features from the speech signals following appropriate audio preprocessing. Noise reduction, down sampling, and silent period removals are some examples of audio preprocessing steps. Different speech features represent different aspects of speech production. For example, energy-based speech features represent respiratory function, whereas pitch-based features represent phonatory function. On the basis of the anticipated speech impairment and speech task characteristics, a set of representative speech features could be extracted. The secondary analysis then encompasses the relationship between speech features and clinical variables through descriptive statistics, statistical comparisons, association mining, and unsupervised data explorations.

In tertiary analysis, research can exploit speech variability within diseases to build predictive models, typically through classification or regression. Insights derived from secondary analysis can be integrated into predictive model building through feature selection and patient clustering. In addition, the impact of confounding factors such as age and gender should be considered when developing predictive models. Algorithm-dependent advanced ML strategies such as feature selection, data augmentation, and transfer learning can be explored to improve model predictions.

Once the solution has reached a satisfactory state, the evaluation of the proposed solutions or approaches is recommended [134]. During evaluations, the hypothesis can be refined to a more precise level based on the data analysis. For example, data analysis might highlight speech impairments in a particular dimension of speech production, such as phonation or articulation, or a particular stage of disease. Therefore, specific hypotheses can lead to detailed insights. Moving forward, internal validations should extend to different dimensions of model performance, and evaluations can include surveys and case studies for clinical utility assessments.

Through a set of comprehensive experiments and evaluation metrics, internal validations should extend beyond prediction performance to include assessments of biases and fairness, reliability, and explainability of predictions to address challenges that persist in speech analysis [28]. With the aid of publicly available datasets or, if feasible, with external cohorts, external validation of predictions can be conducted. Few case studies can be carried out in a clinical environment to present model assessments to clinical experts and obtain their feedback on model usability.

Finally, study results can be presented including experimental results for internal validations, external validations, and feedback from expert observations. We believe that the research outputs will be more competent and thorough and contribute to long-term research directions, extending the short-term results of the specific research scope.