Efficacy of Early Rehabilitation for Severe Coronavirus Disease 2019          Pneumonia: Factor Analysis Using Machine Learning

Mitsuhiko Ikebuchi; Yoichi Ohta; Yukihide Minoda; Akiko Toki; Tamotsu Nakatsuchi; Hidetomi Terai; Hiroaki Nakamura; Ryoichi Kato; Sigeyoshi Nakajima

doi:10.2490/prm.20230027

ABSTRACT

Objectives: Patients with severe coronavirus disease 2019 (COVID-19) who develop pneumonia face the risk of ventilatory muscle disuse in the acute phase, which can result in persistent respiratory impairments in the subacute phase. Although rehabilitation during the acute phase is considered effective, there are limited reports on this topic. Therefore, this study aimed to investigate the effectiveness of acute-phase rehabilitation in patients with severe COVID-19.

Methods: The study included 57 patients (45 men and 12 women; mean age: 63.2±12.1 years) admitted between April and June 2021, all of whom required intubation for respiratory management. Among them, 34 patients underwent acute-phase rehabilitation interventions based on the early goal-directed mobilization protocol. The primary objectives were to assess the occurrence of medical accidents related to acute-phase rehabilitation and evaluate their impact on survival and mobility upon hospital discharge. Statistical techniques and machine learning algorithms were employed for data analysis.

Results: Remarkably, no medical accidents occurred during the acute-phase rehabilitation among the patients. Furthermore, our findings indicated that acute-phase rehabilitation did not influence survival outcomes. However, it did have a positive impact on the mobility of patients upon hospital discharge.

Conclusions: Acute-phase rehabilitation can be safely administered to patients with severe COVID-19 by following an early goal-directed mobilization protocol. This approach may also contribute to improved activities of daily living after discharge.

INTRODUCTION

For patients experiencing pneumonia caused by severe coronavirus disease 2019 (COVID-19) during the acute phase, there are legitimate concerns for the potential disuse of ventilatory muscles as a result of prolonged intubation, respiratory management, initiation of extracorporeal membrane oxygenation (ECMO) or continuous hemodiafiltration (CHDF), and high-dose steroid therapy. Several reports have focused on rehabilitation of COVID-19 patients in the subacute phase.¹^,²^,³^,⁴⁾ These studies suggested that in patients with impaired respiratory function, respiratory rehabilitation is effective for preventing further respiratory function impairment and improving ambulatory ability and cognitive function. Therefore, rehabilitation initiated in the acute phase is considered effective for patients with COVID-19, as is the case for bacterial pneumonia and other respiratory diseases.

However, only a limited number of studies have specifically demonstrated the efficacy of acute-phase rehabilitation in patients with severe COVID-19. In Japan, COVID-19 is designated as a Class II infectious disease, denoting its high-risk nature based on comprehensive consideration of transmissibility and the implications for affected patients. Facilities that can admit patients with this disease are limited, and strict adherence to the use of personal protective equipment (PPE) is necessary during treatment. Because of concerns about potential PPE breaches or failures that lead to secondary infection among healthcare workers, rehabilitation for patients with COVID-19 tends to be avoided. This situation explains the lower number of patients undergoing rehabilitation in the acute phase than in the subacute phase and makes it difficult to enroll an adequate number of patients for a valid multivariate analysis of the effects of rehabilitation.

Therefore, to evaluate the efficacy of acute-phase rehabilitation in patients with severe COVID-19, we employed an analytical technique using machine learning (ML). ML is a computer-based process that relies on patterns and reasoning to efficiently execute a particular task without explicit instructions. Furthermore, the computer system determines the optimal analytical technique rather than having it specified by humans and determines whether the specified outcomes can be predicted from the available factors.

Although ML can determine the optimal analytical technique, the analysis processes that it uses are not identified. The inability to evaluate the degree of impact of each evaluation factor is a disadvantage. Therefore, we introduced a factor analysis technique using a combination of ML and permutation importance (PI) to analyze the degree of impact of each evaluation factor. PI evaluates the impact of each factor by performing simulations after data for the respective factors are altered. Any changes in the calculated results can be attributed to the effect of the factor under investigation. For example, if a simulation performed by changing only Factor A in a certain data group yields significantly different results, it can be determined that Factor A has a significant impact. The objective of this study was to evaluate the efficacy of acute-phase rehabilitation in patients with severe COVID-19 using conventional statistical techniques, ML, and PI.

MATERIALS AND METHODS

Treatment of Severe COVID-19

The Osaka Metropolitan University Hospital admitted patients with severe COVID-19 after allocation by a health center. When patients were brought in, emergency physicians determined whether intubation or respiratory management was required. Patients who required intubation were placed under systemic management. ECMO or CHDF was initiated depending on the patient’s condition. To treat patients with COVID-19, internists specializing in infectious diseases prescribed steroids, antiviral agents, neutralizing antibodies, and immunosuppressants. Muscle relaxants were administered for 48 h. During follow-up, blood gas analyses, chest radiography, and chest computed tomography (CT) were performed as needed to evaluate pneumonia, and the administration of drugs was tapered and withdrawn.

Patients were intubated in the emergency room as required. Although preventive measures against infection were implemented, patients were then transferred to the closed intensive care unit (ICU) to minimize the opportunity for airborne transmission. COVID-19 was treated only in the ICU. Patients who were successfully weaned from a ventilator were not transferred to a general ward but were directly discharged to a post-acute care hospital through the intermediation of a health center.

In the ICU, treatment was provided by a dedicated team that consisted of core members from the original ICU staff and additional physicians, nurses, and physiotherapists under temporary assignment from general clinical departments. This dedicated team was a mixed team without an ingrained culture of early mobilization. In all patients, the posture was changed to the Sim’s position, and they were not placed in the prone position. The nurses did not perform early mobilization. To prevent contact and droplet infections, members of the dedicated team wore full PPE, including eye shields, caps, long-sleeved gowns, N95 masks, and double gloves, when treating patients. Short lectures on the use of full PPE and infection prevention were provided to team members.

Rehabilitation for COVID-19

After patients were transferred to the ICU, they were examined by rehabilitation medicine physicians, and rehabilitation was initiated. Whether rehabilitation was applicable was determined according to the modified Goal Directed Zooming in Rehabilitation Activity (GODZIRA) protocol for COVID-19 (Fig. 1). The original GODZIRA protocol described the early rehabilitation protocol used in the ICU at the Osaka Metropolitan University Hospital. It aimed at early, goal-directed mobilization, and was jointly prepared by the Departments of Rehabilitation Medicine and Emergency Medicine based on previous studies.⁵^,⁶⁾ All patients in this study met the starting criteria during the first examination and did not meet the suspension criteria during their hospitalization.

Fig. 1.

Modified GODZIRA protocol for COVID-19.

Patient Background

During rehabilitation, we emphasized the maintenance of respiratory function and aid for sputum expectoration. Patient exercises, such as breathing exercises, sitting exercises, joint range of motion exercises, muscle-strengthening exercises, standing exercises, or ambulation exercises, were initiated according to the arousal level of the patient. No termination criteria were set for rehabilitation, which was continued until the patients were transferred to another hospital, discharged home, or transitioned to the best supportive care.

Among patients with COVID-19 who were transferred between April and June 2021, a period that mostly included the fourth wave of the COVID-19 outbreak (March 1 to June 20, 2021), this study included 57 patients (45 men and 12 women) with severe conditions who were intubated for respiratory management (see Table 1). The mean age at arrival was 63.2±12.1 years (range 27–84 years). Patients that had been intubated for respiratory management at the sending hospital were included. No patients had started rehabilitation before admission to our hospital. Comorbidities on admission included diabetes mellitus (DM; n=13), dyslipidemia (n=13), renal impairment (n=7), chronic lung disease (CLD; n=15), hypertension (n=29), heart disease (n=9), and cerebrovascular disorder (n=8). ECMO was initiated in 1 patient and CHDF in 8 patients. Chest CT revealed atelectasis in 24 patients on admission. Complications during hospitalization were cerebrovascular disorders in 3 patients and pneumothorax/subcutaneous emphysema in 3 patients.

Thirty-six patients (63.2%; 30 men and 6 women, 59.8±10.4 years) were transferred to another hospital or discharged home (hereafter, surviving patients), including four patients who were transferred while remaining on the ventilator. CHDF was initiated in two patients. During hospitalization, the complications of cerebrovascular disorder, pneumothorax/subcutaneous emphysema, and pulmonary artery thrombosis each occurred in one patient.

Twenty-one patients (36.8%; 15 men and 6 women, 68.9±12.6 years) died. None of the patients were weaned from the ventilator. The age groups of the deceased were as follows: 20–29 years, 0 deaths; 30–39 years, 1 death; 40–49 years, 1 death; 50–59 years, 2 deaths; 60–69 years, 4 deaths; 70–79 years, 10 deaths; 80–89 years, 3 deaths. Among the patients that died, 15 had one or more comorbidities: 4 patients had DM, 7 patients had dyslipidemia, 4 patients had renal impairment, 3 patients had CLD, 11 patients had hypertension, 4 patients had heart disease, and 3 patients had cerebrovascular disorder. ECMO was initiated in 1 patient who died 28 days after admission. CHDF was initiated in 6 patients. During hospitalization, 4 patients developed complications, including cerebrovascular disorder in 2 patients and pneumothorax/subcutaneous emphysema in 2 patients. In the patients who died, mean hospitalization duration was 14.9±8.4 days (4–39 days), and the mean duration of ventilator use was 7.9±6.1 days (3–29 days). Tracheostomy was performed in 16 (28.1%) patients.

Rehabilitation Background

Prior to May 6, 2021, there was no consensus in our hospital regarding the rehabilitation of COVID-19 patients, regardless of disease severity. However, as of May 6, 2021, rehabilitation was initiated for all patients who met the starting criteria. Out of the 7 patients who were hospitalized before May 6, 2021, none had received rehabilitation until that date. In addition, 34 intubated patients received rehabilitation for respiratory management. They were sedated and exhibited muscle relaxation. The mean interval between admission and initiation of rehabilitation was 6.2±7.1 days (1–27 days). In 27 patients who were transferred after May 6, 2021, the mean interval between admission and initiation of rehabilitation was 2.2±1.3 days (1–6 days). Twenty-four patients achieved sitting, 15 achieved standing, and 7 achieved walking. Mobility was evaluated using the Intensive Care Unit Mobility Scale (IMS) score. The mean IMS score at transfer or death was 2.4±3.2 (0–10). Among the patients who died, although 2 patients achieved sitting, no patient achieved standing or walking. In both cases, the number of days from onset to admission was 18 days and neither suffered complications during hospitalization (patients 33 and 34, Table 1).

Table 1. Patient background information

No.	Sex	Age (years)	Height (cm)	Weight (kg)	BMI	Comorbidities	Survival/ death (S/D)	LOH (days)	DOA (days)	DOI (days)	DV (days)	Weaning	T	ECMO and CHDF	CDH	Group	DAR	Rehabilitation achievement	IMS
1	F	79	150	56.4	25.1	HPT, CBV	S	19	3	3	-		T			NR	-		0
2	M	84	170	77.6	26.9	DM	D	13	2	4	-			CHDF		NR	-		0
3	F	83	154	58	24.5	DYS, HPT	D	23	4	4	-		T			NR	-		0
4	M	48	175	72.5	23.7	DYS, HPT	S	7	11	11	6	W				NR	-	Sit	3
5	M	69	163	67	25.2	HPT	S	5	2	3	3	W				NR	-		1
6	M	43	183	89.9	26.8	DYS, HPT, ATL	S	5	7	6	5	W				NR	-		1
7	M	53	185	92.1	26.9	HPT, ATL	S	15	8	8	10	W				NR	-	Sit	1
8	M	73	173	67	22.4		D	7	18	14	-		T			NR	-		0
9	M	54	180	68.2	21.0	HD	S	5	13	13	3	W				NR	-	Sit, stand	1
10	M	63	173	55	18.4	CLD	S	7	6	6	6	W				NR	-		1
11	M	54	155	59.7	24.8	HPT, HD, ATL	S	6	5	5	5	W				NR	-		1
12	M	62	173	104	34.7	DM, HPT, ATL	S	10	5	5	5	W				NR	-		1
13	M	64	170	76	26.3	CLD, ATL	S	8	11	11	7	W				NR	-		1
14	M	72	178	76	24.0	DYS, CLD, HPT	D	17	5	5	-		T			NR	-		0
15	M	60	172	90	30.4	CLD, HPT	D	17	5	5	-		T			NR	-		0
16	M	69	165	57.8	21.2	HPT, ATL	D	16	5	5	-		T			NR	-		0
17	M	72	172	64.3	21.7	ATL	S	8	11	11	7	W				NR	-	Sit	3
18	M	57	178	71	22.4	CLD, HPT, ATL	S	5	17	17	4	W				NR	-		1
19	M	27	181	95	29.0		S	9	8	8	4	W				NR	-	Sit, stand, walk	9
20	M	67	163	50.8	19.1	DM, HPT	S	4	11	11	3	W				NR	-		1
21	M	66	165	64.7	23.8	HPT, ATL	S	8	6	6	4	W				NR	-	Sit, stand, walk	8
22	M	51	164	64.3	23.9		S	9	5	5	4	W				NR	-		1
23	M	66	165	66	24.2		S	6	6	6	5	W				NR	-	Sit	3
24	M	59	167	68	24.4	CLD	S	19	4	4	12	W				R	19	Sit, stand	4
25	M	56	173	80.6	26.9	DM, DYS, HPT, ATL	S	34	11	11	24	W		CHDF	CBV	R	23		0
26	M	55	170	89.3	30.9	DM, HD, ATL	D	14	12	12	-			CHDF		R	14		0
27	F	77	153	55.9	23.9	DYS, CLD, HPT, ATL	S	19	6	6	-		T			R	10		1
28	M	77	156	60	24.7	DYS, HPT	D	13	6	6	-					R	9		0
29	M	32	163	86.8	32.7	ATL	D	28	9	9	-			ECMO	CBV	R	25		0
30	M	54	167	100	35.9	DM, HPT, HD	S	12	9	9	11	W		CHDF		R	9		1
31	F	75	153	57.6	24.6	DYS, CLD	D	16	8	8	-	-	T			R	9		0
32	M	46	175	75.5	24.7	CLD, ATL	S	8	12	11	6	W				R	7	Sit, stand	9
33	M	56	174	70.3	23.2	ATL	D	14	18	18	-			CHDF		R	5	Sit	0
34	M	77	161	72	27.8	DM, DYS, RI, HPT, ATL	D	24	18	18	-		T			R	5	Sit	0
35	M	57	160	57.3	22.4	CLD, ATL	S	6	12	12	5	W				R	2	Sit	3
36	M	71	168.9	61.1	21.4	ATL	D	35	11	12	-		T		CBV	R	2		0
37	M	50	169	84.1	29.4	DYS, HPT, HD, ATL	S	39	9	9	21	W				R	27	Sit, stand	4
38	F	74	156	64.8	26.6	HPT, HD, ATL	D	16	5	5	-		T			R	2		0
39	F	66	154	43	18.1	DYS, RI, HPT, HD	D	19	19	19	-			CHDF		R	3		0
40	M	68	185	91	26.6	DM, HPT, ATL	S	9	12	8	-					R	4		1
41	M	68	168	68	24.1	HPT, ATL	D	10	15	15	-				P/E	R	3		0
42	F	71	158	55	22.0	DYS, HPT	D	19	5	6	-		T		P/E	R	3		0
43	F	58	163	84	31.6	DM, CLD, HPT	S	35	9	9	29	W	T		P/E	R	1	Sit	6
44	M	77	164	67	24.9	RI, HPT	D	16	5	5	-		T	CHDF		R	3		0
45	M	46	176	78	25.2		D	27	9	9	-		T	CHDF		R	2		0
46	M	72	174	69.3	22.9	CLD, HPT	S	15	3	3	7	W				R	6	Sit, stand, walk	10
47	F	52	163	68.7	25.9	ATL	S	16	7	7	9	W				R	1	Sit, stand, walk	9
48	F	71	160.2	57.2	22.3	RI	S	21	12	12	15	W	T			R	3	Sit, stand	4
49	M	67	173	90.7	30.3	RI, CLD, ATL	S	21	2	2	-					R	1	Sit, stand	4
50	M	73	164	53.1	19.7	CLD	S	8	14	14	5	W				R	1	Sit, stand, walk	9
51	M	57	166	67	24.3		S	21	8	8	5	W				R	1	Sit, stand, walk	10
52	F	61	165	54.3	19.9	DM	S	14	2	2	5	W				R	1	Sit, stand	6
53	M	55	176	75	24.2	CLD, HPT	S	8	14	14	6	W				R	1	Sit, stand	5
54	M	70	171.1	78.85	26.9	DM, RI, HPT, HD	S	9	9	9	6	W				R	2	Sit	3
55	F	79	160	70.3	27.5		D	25	4	4	-					R	2		0
56	M	82	165	56.6	20.8	DM, RI, HD	D	19	12	12	-					R	1		0
57	M	55	172	67.7	22.9	DM, DYS	S	12	9	9	7	W				R	4	Sit, stand, walk	10
Mean		63.2	167.5	70.6	25.0			14.9	8.7	8.6	7.9						6.2		2.4
SD		12.1	8.4	13.4	3.7			8.4	4.4	4.2	6.1						7.1		3.2

BMI, body mass index; S; survival (transferred to another hospital or discharged home); D, death; LOH, length of hospitalization; DOA, duration from onset to admission; DOI, duration from onset to intubation; DV, duration on ventilator; T, tracheostomy; ECMO, extracorporeal membrane oxygenation; CHDF, continuous hemodiafiltration; CDH, complications during hospitalization; DAR, days from admission to start of rehabilitation; IMS, Intensive Care Unit Mobility Scale; SD, standard deviation; HD, heart disease; HPT, hypertension; RI, renal impairment; ATL, atelectasis; DYS, dyslipidemia; CBV, cerebrovascular disorder; W, weaned; P/E, pneumothorax/emphysema; R, Rehabilitation group; NR, Non-rehabilitation group

The patients were divided into a non-rehabilitation group (NR group) and a rehabilitation group (R group), and the two groups were compared for 22 study items, including indications for rehabilitation. ML was used to perform a factor analysis with these study items as evaluation factors, and the impact on survival/death was evaluated with “transfer to another hospital/discharge to home or death” as an objective variable. The 22 study items included risk factors for severe COVID-19 [sex, age, height, body weight, and body mass index (BMI)], comorbidities (DM, dyslipidemia, renal impairment, CLD, hypertension, heart disease, cerebrovascular disorder, and atelectasis), days from onset to admission, days from onset to intubation, weaning, tracheostomy, initiation of ECMO and CHDF, complications during hospitalization (cerebrovascular disorder and pneumothorax/emphysema), and indications for rehabilitation. Surviving patients in the NR and R groups were compared in the same manner, and the impact on mobility upon hospital discharge was evaluated using the IMS score as the objective variable. Given that ECMO was not initiated in any surviving patients, the impact on survival/death was evaluated using 21 study items.

Data Analysis

For the two-group comparison, an unpaired t-test and Fisher’s exact test were performed, and P<0.05 was considered statistically significant. The statistical calculation software R for Mac OS X Cocoa GUI version 3.6.2 (R Core Team)⁷⁾ was used for analysis.

Multiple-factor analysis was performed using Twink analysis software (Department of Core Informatics, Graduate School of Informatics, Osaka Metropolitan University). The factors were analyzed using ML with hierarchical neural networks for structure, general backpropagation neural networks for learning methods, and PI for factor evaluation. General backpropagation neural networks are also used in the statistical calculation software R. As a result of learning through an exclusive OR gate, which is a textbook accuracy verification method in information engineering, learning was performed with an error rate of 0.001%.

The moderating variables were “survival/death” and “mobility (IMS score).” The explanatory variables were the aforementioned study items. The hidden layer consisted of one layer, and the number of items in the hidden layer was set to seven. In ML, both moderating and explanatory variables must be normalized on a scale of 0 to 1. For nominal variables, such as survival/death, sex, and the presence or absence of complications, a value of 0 was assigned to “absent (or dead or female),” and a value of 1 was assigned to “present (or alive or male).” Continuous variables such as age, BMI, and IMS score were normalized on a scale of 0 to 1 based on the mean and standard deviation.

The accuracy of the ML results was expressed on a receiver operating characteristic (ROC) curve. In this study, an accurate learning model was considered to be achieved when the area under the curve (AUC) was 0.9 or higher. PI indicates the degree of impact when a certain factor is changed in a positive or negative direction. To calculate PI, the factors were changed within the available data. An output value of 0 to +1 indicated a positive impact on the moderating variable, and an output value of −1 to 0 indicated a negative impact. The magnitude of the numerical value indicated the degree of impact (see Fig. 2). Referring to Pearson’s correlation coefficient, we defined a PI value of 1–0.75 as strong impact, 0.75–0.5 as moderate impact, 0.5–0.25 as weak impact, and less than 0.25 as no impact. In this study, we considered that factors had an impact when the magnitude of the PI value was 0.25 or higher.

Fig. 2.

Examples of impact indicated by permutation importance: factor A, positive impact; factor B, negative impact; factors C and D, constant impact; factor E, no impact.

Ethical Considerations

Written informed consent could not be obtained because of the constraints imposed by the retrospective design. Participants could withdraw from this study at any time using an opt-out procedure. This study was reviewed and approved by the ethics committee of Osaka Metropolitan University Graduate School of Medicine (2021–086).

RESULTS

Table 2 shows the results of the two-group comparison between the NR and R groups for all the patients. There were 57 patients in total (23 in the NR group and 34 in the R group). We surveyed the risk factors for severe COVID-19. A statistically significant difference was only observed in the number of patients with renal impairment. The days from onset to admission and days from onset to intubation were not significantly different between the two groups. The duration of hospitalization and ventilation was significantly longer in the R group than in the NR group. Seventeen patients in the NR group and 19 patients in the R group survived, showing no statistically significant difference between the two groups. Weaning, tracheostomy, ECMO, and CHDF were not significantly different between the two groups. Complications during hospitalization were not significantly different between the two groups. Failure of the ventilator closed circuit, breach of PPE during rehabilitation sessions, and close contact/secondary infection for treating personnel did not occur in any case. The sitting, standing, walking, and IMS score at discharge/transfer or death were not significantly different between the two groups.

Table 2. Comparison of NR and R groups for all patients

			Non-rehabilitation (NR) group (n=23)	Rehabilitation (R) group (n=34)	P value
Risk factors for severe COVID-19
	Sex (M/F)		21/2	24/10	N.S.
	Age (years)		62.4±12.9	63.6±11.5	N.S.
	Height (cm)		169.9±9.1	166.0±7.4	N.S.
	Body weight (kg)		71.4±13.9	69.9±13.0	N.S.
	BMI		24.6±3.5	25.3±3.8	N.S.
	Comorbidities:
		Diabetes mellitus	3	10	N.S.
		Dyslipidemia	4	9	N.S.
		Renal impairment	0	7	<0.05
		Chronic lung disease	5	10	N.S.
		Hypertension	14	15	N.S.
		Heart disease	2	7	N.S.
		Cerebrovascular disorder	4	4	N.S.
		Atelectasis	9	15	N.S.
Days from onset to admission			7.6±4.3	9.4±4.4	N.S.
Days from onset to intubation			7.5±3.8	9.3±4.4	N.S.
Hospitalization days			10.0±5.2	18.3±8.4	<0.05
Ventilator days			5.1±1.8	10.8±7.4	<0.05
Survival/death			17/6	19/15	N.S.
Weaning			16	16	N.S.
Tracheostomy			6	10	N.S.
ECMO			0	1	N.S.
CHDF			1	7	N.S.
Complications during hospitalization
	Cerebrovascular disorder		0	3	N.S.
	Pneumothorax/emphysema		0	3	N.S.
Rehabilitation
	Sitting		7	17	N.S.
	Standing		3	12	N.S.
	Walking		2	5	N.S.
	IMS (at transfer or death)		1.6±2.3	2.9±3.6	N.S.

Data given as number or mean±SD.

N.S., not significant.

Table 3 shows the results of the two-group comparison of the surviving patients between the NR and R groups. Thirty-six patients survived: 17 in the NR group and 19 in the R group. We surveyed the risk factors for severe COVID-19. A statistically significant difference was only observed in the number of patients with renal impairment. The duration from onset to admission and from onset to intubation was not significantly different between the two groups. The duration of hospitalization and ventilation was significantly longer in the R group than in the NR group. Weaning, tracheostomy, ECMO, and CHDF were not significantly different between the two groups. Complications during hospitalization were not significantly different between the two groups. More patients in the R group acquired sitting, standing, and walking abilities than those in the NR group, with significant differences between the two groups. The IMS score at discharge/transfer was significantly higher in the R group than in the NR group.

Table 3. Comparison of NR and R groups for surviving patients

			Non-rehabilitation (NR) group (n=17)	Rehabilitation (R) group (n=19)	P value
Risk factors for severe COVID-19
	Sex (M/F)		16/1	14/5	N.S.
	Age		58. ±12.0	60.9±8.6	N.S.
	Height (cm)		170.3±9.5	168.2±7.1	N.S.
	Body weight (kg)		71.6±14.7	72.5±13.3	N.S.
	BMI		24.5±3.7	25.5±3.9	N.S.
	Comorbidities:
		Diabetes mellitus	2	7	N.S.
		Dyslipidemia	2	4	N.S.
		Renal impairment	0	3	<0.05
		Chronic lung disease	3	9	N.S.
		Hypertension	10	8	N.S.
		Heart disease	2	3	N.S.
		Cerebrovascular disorder	3	2	N.S.
		Atelectasis	8	8	N.S.
Days from onset to admission			7.9±3.8	8.6±3.7	N.S.
Days from onset to intubation			7.9±3.7	8.4±3.6	N.S.
Hospitalization days			8.0±3.7	17.2±9.5	<0.05
Ventilator days			5.1±1.8	10.8±7.4	<0.05
Weaning			16	16	N.S.
Tracheostomy			1	3	N.S.
ECMO			0	0	N.S.
CHDF			0	2	N.S.
Complications during hospitalization
	Cerebrovascular disorder		0	1	N.S.
	Pneumothorax/emphysema		0	1	N.S.
Rehabilitation
	Sitting		7	15	<0.05
	Standing		3	12	<0.05
	Walking		2	5	N.S.
	IMS (at transfer or death)		2.2±2.5	5.2±3.3	<0.05

Data given as number or mean±SD.

N.S., not significant.

Figure 3 shows the results of factor analysis performed using ML. When the moderating variable was survival/death, the ROC curve yielded a sensitivity of 100%, a specificity of 100%, and an AUC of 1.00, indicating that outcomes could be accurately predicted from the 22 explanatory variables. When the moderating variable was mobility (IMS score), the ROC curve yielded a sensitivity of 100%, a specificity of 100%, and an AUC of 1.00, indicating that outcomes could be accurately predicted from the 21 explanatory variables.

Fig. 3.

ROC curves for factor analysis using ML: left, survival/death as moderating variable; right, mobility (IMS score) as moderating variable.

Figures 4 and 5 show the results of the multiple-factor analysis performed using PI. When the moderating variable was survival, the absence of dyslipidemia and CLD among comorbidities, as well as successful weaning, may have had a positive impact on survival, whereas the absence of cerebrovascular disorder occurring as a complication during hospitalization and failed weaning may have had a negative impact. These results indicate that rehabilitation did not significantly affect survival or death.

Fig. 4.

Multiple factor impact analysis by PI with survival/death as the moderating variable.

Fig. 5.

Multiple factor impact analysis by PI with mobility (IMS score) as the moderating variable.

When considering the IMS score as the moderating variable, the absence of renal impairment, CLD, heart disease, and cerebrovascular disorders among comorbidities, the absence of cerebrovascular disorder complication during hospitalization, and the presence of coexisting pulmonary emphysema/subcutaneous emphysema had a positive impact on the IMS score, whereas the presence of CLD, heart disease, and cerebrovascular disorder as comorbidities, the presence of cerebrovascular disorder, and the absence of pulmonary emphysema/subcutaneous emphysema as complications during hospitalization had a negative impact on the IMS score. Rehabilitation had a positive impact on the IMS score when interventions were performed but had a negative impact when they were not performed.

DISCUSSION

The timing of rehabilitation for patients with COVID-19 is controversial in terms of transmissibility and the risk of progression to severe conditions. We applied the GODZIRA protocol, which is used for early interventions in the general ICU of our hospital. Liu et al.⁸⁾ also developed a protocol to allow for early mobilization, and they reported that it allowed them to perform early mobilization without causing complications associated with rehabilitation. In this study, the average time from admission to the start of rehabilitation was 6.2±7.1 days (1–27 days) for all patients. However, when the analysis was limited to patients who were transferred after the indications for rehabilitation were started, the average time dropped to 2.2±1.3 days (1–6 days). The GODZIRA protocol allowed early rehabilitation to be started smoothly for patients with COVID-19.

When rehabilitation was performed by staff in full PPE who were fully versed in procedures and the layout of the COVID-19 isolation ward, they were able to respond not only to sudden changes in patients’ conditions but could also perform interventions without causing any problems, such as breach of PPE and close contact/secondary infection for treating personnel. Early rehabilitation performed in patients with severe COVID-19 appears to be safe with sufficient preparation. In the chronic phase, lung function may remain impaired even after recovery in patients with COVID-19.⁹^,¹⁰^,¹¹⁾ Daynes et al.¹²⁾ applied respiratory rehabilitation to 30 patients with COVID-19 125 days after infection and reported an improvement in exercise capacity and respiratory function. Therefore, rehabilitation initiated in the acute phase may be effective. The Stanford Hall consensus statement recommends the consideration of a holistic approach because several patients with COVID-19 who require rehabilitation have to coexist with post-intensive care syndrome, as is the case with other ICU patients.¹³⁾ Ceravolo et al.¹⁴⁾ also proposed, based on their systematic review, that early rehabilitation is applicable to hospitalized patients with COVID-19. However, among reports of rehabilitation initiated on admission to the ICU, Ozyemisci Taskiran et al.¹⁵⁾ reported that rehabilitation was not effective for the recovery of muscle strength when 18 patients with COVID-19, who were admitted to the ICU and underwent rehabilitation in the acute phase, were compared with 17 patients who did not undergo rehabilitation.

Our results showed that acute-phase rehabilitation interventions did not greatly affect survival or death. Regarding mobility upon hospital discharge in surviving patients, the two-group comparison showed that the IMS score was significantly higher in the R group than in the NR group, and the analysis using ML also showed that rehabilitation had a positive impact. This study suggests that acute-phase rehabilitation for patients with severe COVID-19 does not greatly affect prognosis but may improve activities of daily living after discharge. Reports on the effects of early mobilization in patients admitted to the ICU indicated that it shortens the length of ICU stay and improves mobility after discharge from the ICU.⁵^,⁶^,⁷^,⁸^,⁹^,¹⁰^,¹¹^,¹²^,¹³^,¹⁴^,¹⁵^,¹⁶^,¹⁷⁾ The results of this study appear to be comparable to those described in these reports.

Early mobilization of patients in the ICU reportedly shortens the length of ICU stay.⁵^,¹⁷⁾ However, in this study, the duration of hospitalization was significantly longer in the R group than in the NR group. This seems to be attributable to the social circumstances surrounding COVID-19 in Japan. During the second half of the fourth wave of the COVID-19 outbreak, when acute-phase rehabilitation intervention was initiated at our hospital, all beds for patients with post-acute COVID-19 were occupied in all municipalities. Consequently, many patients could not be transferred from our hospital even after surviving the acute phase. This may have been reflected in the results of this study.

Regarding the impact of other factors on survival/death, ML analysis showed that dyslipidemia negatively affected survival in addition to CLD, atelectasis, and cerebrovascular disorder as comorbidities on admission, whereas successful weaning positively affected survival. For the impact on the IMS score, ML analysis suggested that renal impairment, CLD, heart disease, and cerebrovascular disorder upon admission, as well as cerebrovascular disorder during hospitalization, might negatively affect the IMS score, whereas pulmonary emphysema/subcutaneous emphysema might positively affect the IMS score.

The COVID-19 Registry Japan (COVIREGI-JP) is a registry of patients with COVID-19 in Japan.¹⁸⁾ Tanaka et al.¹⁹⁾ studied 1529 patients who were registered in the COVIREGI-JP and were intubated for respiratory management. There were 401 deaths (306 men, 95 women), accounting for 26.3% of the total. The mean age of survivors was 64.3±12.3 years and the average age of those who died was 73.4±9.9 years. They reported that age is associated with the mortality rate in patients intubated for respiratory management.¹⁹⁾ Asai et al.²⁰⁾ studied 4701 patients and reported that the mortality rate was higher for older patients, the male sex is an important risk factor for survival/death, and the coexistence of connective tissue disease and renal disorder, as well as the concurrent use of dialysis, are risk factors for survival/death.²⁰⁾

In the current study, age was not a factor that greatly affected survival/death according to either the two-group comparison or analysis using ML. This result may have been caused by possible bias in the age distribution of the patients included in this study. The mean age of the patients in the COVIREGI-JP study was approximately 10 years older than that of the patients in the current study.¹⁹⁾ Selection bias may have occurred in the current study because the patients were allocated by a health center in which administrative issues, social circumstances, and other factors may have influenced patient allocation. The prevalence of comorbidities and incidence of complications occurring during hospitalization were also assumed to be associated with older age. Similar to the effect of age, our results suggested that comorbidities were not factors that greatly affected survival/death. These factors may have been also indirectly influenced by the policies of the heath center.

The presence of atelectasis positively affected survival, suggesting that exacerbation caused by atelectasis could be easily resolved if the factors leading to atelectasis were removed. However, because data on the course of respiratory management and respiratory function, such as blood gas and oxygen saturation levels, were not included in this study, further investigation is necessary. Regarding mobility upon hospital discharge, analysis using ML and PI showed that BMI and comorbidities, except for dyslipidemia and hypertension, negatively affected the IMS score. Given that these underlying diseases are considered to be factors for the progression to severe COVID-19, it seems natural that they are factors that inhibit recovery. The data collected on other underlying diseases in this study may also have been affected by bias, as mentioned earlier. As data accumulate, future studies may demonstrate that other comorbidities negatively affect IMS score. The presence of coexisting pulmonary or subcutaneous emphysema had a positive impact on the IMS score. This suggests that prioritizing oxygenation may contribute to the recovery of activity, even if lung damage occurs because of positive pressure. This study also implies the predictive potential for ML and PI to identify relevant factors; however, further research is required for more conclusive evidence.

Clinical research on rehabilitation medicine requires the evaluation of more factors than in studies in other fields. Hence, it is difficult to enroll the necessary number of participants to ensure the reliability of the multivariate analysis. Therefore, we introduced analytical techniques using ML and PI. Conventional statistical methods (e.g., multivariate analysis) sort explanatory variables using mathematical formulas. In clinical data, outliers are unavoidable. Therefore, to remove the influence of outliers, it is necessary to strictly define the number of factors and data. In ML, the program sorts explanatory variables and because no formula is used, outliers are suppressed. As a result, restrictions on the number of factors and data are not imposed. The essence of PI is to input data into the program, which then outputs the results. Whether undertaking ML or PI, if there is no bias in the data, factor analysis is possible even with a small number of data groups. Altmann et al.²¹⁾ analyzed life science sample data with ML with random forest and PI to calculate P values and evaluate factors. They demonstrated the usefulness of these analytical techniques.²¹⁾

In this study, we introduced statistical techniques using ML with a hierarchical neural network and PI. As a result, we were able to perform ROC curve analysis with high accuracy, with an AUC of 1.00, and identify several affecting factors. We demonstrated the usefulness of analytical techniques using ML and PI in clinical research of rehabilitation medicine.

ML is not strictly restricted by sample size but cannot determine the appropriateness of the factors. Therefore, there is more of a need to identify the appropriate factors for ML than for conventional statistical techniques. In this study, we used PI to perform a factor analysis. Factor analysis using PI was performed with one factor and moderating variable. When there are multiple factors (e.g., both DM and hypertension as comorbidities), a separate factor that combines multiple factors must be prepared. Because the analysis of all combinations of factors would require infinite computational power, it is necessary to prepare a medically appropriate combination of factors. Given these difficulties, the use of analytical techniques using hierarchical neural networks and PI, which were used in this study, is not absolute. Further studies are needed to explore analytical techniques that are more suitable for studies of rehabilitation medicine.

This study had several limitations. The first limitation was the sample size. Although our hospital admits a relatively large number of patients with severe COVID-19, a much larger number of patients is needed to perform valid statistical analyses. In particular, it is well known that the COVID-19 virus rapidly mutates, and the characteristics of each outbreak vary. In addition to the accumulation of cases over time, the acquisition of cases through multicenter cooperation and other means should be discussed. The small size of the sample is a limitation for the evaluation of ML models in this study. We did not perform cross-validation, which involves dividing the dataset into two groups for machine learning and cross-validation purposes. Given the limited number of cases (n=57), such division could introduce bias and result in overfitting of the ML model. Therefore, cross-validation was intentionally omitted because we believed it could potentially exacerbate overfitting issues rather than resolve them. The second limitation was possible bias caused by the involvement of a health center in the allocation of patients. During the COVID-19 outbreak, healthcare resources dwindled, and the allocation of patients by a health center became necessary. This practice appears to be an inevitable bias in the clinical research on COVID-19. Because the number of facilities that admit patients with severe COVID-19 is limited, the acquisition of cases through multicenter cooperation may be helpful for mitigating this bias. Third, we did not follow-up with patients after they were transferred. Because a health center also determined the hospital to which patients were transferred, they were transferred to a broad range of hospitals. Consequently, it is difficult for such a follow-up study to be reviewed and approved by the ethics committees of the hospitals concerned. Therefore, we decided not to follow-up on the patients in this study. Fourth, this study was not a randomized controlled trial. The duration of hospitalization was significantly longer in the R group than in the NR group. Therefore, the IMS score may have improved naturally because of the longer duration of hospitalization. Follow-up studies regarding this point may provide further insights. Other limitations include the exclusion of data that may exhibit characteristic changes over time, such as vital signs, hematologic test results, and drug doses. The input of such data into ML requires additional processing. These factors need to be incorporated into future studies.

CONCLUSION

In this study, we investigated the efficacy of early rehabilitation in patients with severe COVID-19. Both statistical analysis and analysis using ML suggested that rehabilitation might not affect survival/death but can yield a positive impact on mobility upon hospital discharge. Although our results regarding other factors were different from those based on the Japan COVID-19 Registry, we suspect that some bias was introduced through our association with the health center. Therefore, further investigations are required.

CONFLICTS OF INTEREST

The authors declare no conflict of interest.

REFERENCES

Corresponding author

Register with J-STAGE for free!