Table 4 Studies evaluating the type and number of features required for machine learning algorithms to predict the outcomes of TSA

Kumar et al., (2021) [31]

19 and 291

-Full feature group (291) includes:

Demographic data, diagnoses, comorbidities, implant type, preoperative ROM, preoperative radiographic findings, and preoperative PROMs (ASES, SPADI, SST, UCLA, and Constant metrics)

-Minimal feature group (19) includes:

Age, weight, height, sex, Previous shoulder surgery, surgery on dominant arm, diagnosis, comorbidities, ROM, Global Shoulder Function score, VAS score, Pain at worse, pain when lying on the side, pain when touching back of neck, and pain when pushing with effective arm

ASES, Constant score, global shoulder function score, VAS pain scores, active abduction, forward elevation, and external rotation

A comparison of MAEs for the full and minimal models shows that each model had comparable prediction accuracy for each outcome measure. When the minimal model was augmented with data on implant size and/or type, as well as measurements of native glenoid anatomy, only minor improvements in MAEs were seen for each outcome measure

Kumar et al., (2021) [32]

291

Demographic data, diagnoses, comorbidities, implant type, preoperative ROM, preoperative radiographic findings, and preoperative PROMs (ASES, SPADI, SST, UCLA, and Constant metrics)

N/A

The predictive value of the questions in the UCLA score exceeded that of the Constant questions, while the Constant questions were more predictive than the ASES questions. In addition, the preoperative SPADI score was more predictive than the preoperative ASES, Constant, and UCLA scores. Ultimately, we determined that subjective self-assessments of pain, as well as objective measurements of active range of motion and strength, were the most predictive types of preoperative input questions

Kumar et al., (2022) [29]

19 and 291

-Full feature group (291) includes:

Demographic data, diagnoses, comorbidities, implant type, preoperative ROM, preoperative radiographic findings, and preoperative PROMs (ASES, SPADI, SST, UCLA, and Constant metrics)

-Minimal feature group (19) includes:

Age, weight, height, sex, Previous shoulder surgery, surgery on dominant arm, diagnosis, comorbidities, ROM, Global Shoulder Function score, VAS score, Pain at worse, pain when lying on the side, pain when touching back of neck, and pain when pushing with effective arm

-Minimal feature set + implant data includes:

All the features from the Minimal feature group plus implant size/type data, and computed tomographic and radiographic-based measurements of native glenoid version and inclination

Internal Rotation

At each prediction time point, the degree of prediction accuracy across the three major model input categories was comparable between the full feature model and the minimal feature with and without implant/imaging data

The preoperative composite ROM score was found as the most relevant feature driving each minimal feature set, whereas the follow-up length was the most meaningful factor driving the Full feature group, with composite ROM score being the third most important feature

Kumar et al., (2022) [30]

19 and 291

-Full feature group (291) includes:

Demographic data, diagnoses, comorbidities, implant type, preoperative ROM, preoperative radiographic findings, and preoperative PROMs (ASES, SPADI, SST, UCLA, and Constant metrics)

-Minimal feature group (19) includes:

Age, weight, height, sex, Previous shoulder surgery, surgery on dominant arm, diagnosis, comorbidities, ROM, Global Shoulder Function score, VAS score, Pain at worse, pain when lying on the side, pain when touching back of neck, and pain when pushing with effective arm

-Minimal feature set + implant data includes:

All the features from the Minimal feature group plus implant size/type data, and computed tomographic and radiographic-based measurements of native glenoid version and inclination

SAS score, ASES score, Constant score

SAS score was the most predictive/accurate variable to predict aTSA and rTSA outcomes for all 3 machine learning techniques followed by the Constant score and finally the ASES score. For all the outcomes, follow-up duration was the most important feature for the Full feature group, while composite ROM was the most important feature for the minimal feature group

McLendon et al., (2021) [35]

N/A

-Model 1: using of all baseline variables

-Model 2: omitting morphological variables

-Model 3: omitting ASES variables

ASES

Latent factors and morphological variables had most accurate predictions when combined which suggests that both structural pathology and patient perceptions are important for achieving the best results/predictions

Polce et al., (2020) [36]

16

Age, BMI, sex, insurance status, preoperative duration of shoulder-related symptoms > 2 years (yes or no), smoking status, history of ipsilateral shoulder surgery, diabetes mellitus or HTN, preoperative physical activity, humeral component fit, diagnosis, procedure (aTSA or rTSA), ASES, SANE, and subjective Constant-Murley score

Patient satisfaction

Age, insurance status, smoking status, BMI, diabetes mellitus, preoperative activity, preoperative duration of symptoms, diagnosis, procedure, and baseline SANE score were the 10 predictive factors revealed by RFE and cross-validation during model training

The baseline SANE score, exercise and activity, insurance status, diagnosis, and preoperative duration of symptoms were the five most predictive variables that went into the SVM model when they were averaged across all patients