Skip to main content
. 2022 May 3;12:7166. doi: 10.1038/s41598-022-11012-2

Table 4.

Comparison of top-50 important features between suboptimal models and the best performing one.Importance of some comorbidities tend to be overestimated by suboptimal models, while the importance of peripheral vascular is underestimated in AutoInt and LSTM. Suboptimal models underestimate various labevent and chartevent features.

Model Extra (identified in model but not in IMVLSTM) Lacking (identified in IMVLSTM but not in model)
AutoInt

88. systolic_blood_pressure_abp_mean

90. body_temperature

134. congestive_heart_failure

135. cardiac_arrhythmias

140. paralysis

152.solid_tumor

153. rheumatoid_arthritis

160.alcohol_abuse

163. depression

31. HEMOGLOBIN

33. MCH

44. INR(PT)

45. PT

57. LACTATE

76. O2Flow

82. SpO2DesatLimit

91. pao2

138. peripheral_vascular

LSTM

32. MCHC

135. cardiac_arrhythmias

136. valvular_disease

143. diabetes_uncomplicated

146. renal_failure

39. CREATININE

57. LACTATE

82. SpO2DesatLimit

131. language

138. peripheral_vascular

TCN

77. Glucosefingerstick

134. congestive_heart_failure

136. valvular_disease

151. metastatic_cancer

152. solid_tumor

39. CREATININE

57. LACTATE

82. SpO2DesatLimit

93. urinary_output_sum

Transformer

135. cardiac_arrhythmias

136. valvular_disease

143. diabetes_uncomplicated

152. solid_tumor

45. PT

57. LACTATE

82. SpO2DesatLimit

124. HEM